luke robinson thesis revision · 2019-08-21 · figure 24: damage feature statistical performance...
TRANSCRIPT
UNIVERSITY OF CALIFORNIA, SAN DIEGO
A Performance Comparison of Condition Based Monitoring Damage Features Used in
Rotating Machines under Variable Conditions
A Thesis submitted in partial satisfaction of the requirements for the degree Master of Science
in
Structural Engineering: Specialization in Health Monitoring, Prognosis and Validated Simulations
by
Luke Thomas Robinson
Committee in charge:
Professor Michael Todd, Chair Professor Chuck Farrar Professor Bill Hodgkiss
2013
Copyright
Luke Thomas Robinson, 2013
All rights reserved.
iii
The Thesis of Luke Thomas Robinson is approved, and it is acceptable
in quality and form for publication on microfilm and electronically:
Chair
University of California, San Diego
2013
iv
TABLE OF CONTENTS
Signature Page ............................................................................................................................... iii
Table of Contents ........................................................................................................................... iv
List of Abbreviations ..................................................................................................................... vi
List of Figures ............................................................................................................................... vii
List of Tables ................................................................................................................................ xii
Acknowledgements ..................................................................................................................... xiii
Vita .............................................................................................................................................. xiv
Abstract ......................................................................................................................................... xv
Chapter 1 Introduction to Structural Health Monitoring and Condition Based Monitoring of Rotating Machines .......................................................................................................................... 1
1.1. Introduction................................................................................................................ 1
1.2. Structural Health Monitoring Process ........................................................................ 2
1.2.1. Operational Evaluation ................................................................................. 4 1.2.2. Data Acquisition ........................................................................................... 5 1.2.3. Feature Selection .......................................................................................... 6 1.2.4. Statistical Model Development ..................................................................... 7
1.3. Condition Based Monitoring of Rotating Machines .................................................. 9
1.3.1. Operational Evaluation ................................................................................. 9 1.3.2. Data Acquisition ......................................................................................... 10 1.3.3. Feature Selection ........................................................................................ 11 1.3.4. Statistical Model Development .................................................................. 12
1.4. Scope of Thesis ........................................................................................................ 13
Chapter 2 Technical Background ................................................................................................. 15
2.1. Introduction.............................................................................................................. 15
2.2. Digital Signal Processing Techniques ..................................................................... 15
2.2.1. Fourier Transform & Power Spectral Density ............................................ 16 2.2.2. Time-Frequency Analysis ........................................................................... 18 2.2.3. Signal Envelope and the Hilbert Transform ............................................... 20
2.3. Condition Based Monitoring Feature Extraction ..................................................... 20
2.3.1. Angular Resampling ................................................................................... 21 2.3.2. Time Synchronous Averaging .................................................................... 23 2.3.3. Discrete Random Separation ...................................................................... 24 2.3.4. The Difference, Residual and Band-pass Mesh Signal ............................... 25
v
2.3.5. Traditional Damage Features ...................................................................... 26 2.3.6. Hölder Exponent and the FMH Feature ...................................................... 29
Chapter 3 Experimental Setup & Feature Extraction Results ....................................................... 33
3.1. Damage Types and Descriptions ............................................................................. 33
3.2. Experimental Setup .................................................................................................. 33
3.2.1. Testbed Description and Evaluation ........................................................... 33 3.2.2. Data Acquisition ......................................................................................... 37 3.2.3. Description of System States .............................................................. 41 3.2.4. Testing Procedures...................................................................................... 49 3.2.5. Sources of Variability ................................................................................. 52
3.3. Statistical Modeling and Detection Performance .................................................... 53
3.3.1. Hypothesis Testing & Receiver Operating Characteristic Curves .............. 53
3.4. Testing Results ......................................................................................................... 59
3.4.1. Worn Tooth Gearbox Damage Preliminary Results ................................... 61 3.4.2. Main Shaft Ball Bearing Damage Preliminary Results ............................. 72 3.4.3. Bearing Damage Detection and Varying Belt Tension .............................. 82 3.4.4. Worn tooth Damage and Varying Torsional Load ..................................... 91
Chapter 4 Conclusion & Future Work .......................................................................................... 98
4.1. Conclusion ............................................................................................................... 98
4.2. Recommendations for Future Work ........................................................................ 99
References .................................................................................................................................. 101
vi
LIST OF ABBREVIATIONS
AR Autoregressive
ARS Angular Resampled Signal
CBM Condition Based Monitoring
CWT Continuous Wavelet Transform
DAQ Data Acquisition
DRS Discrete/Random Separation
DSP Digital Signal Processing
DVWD Discrete Wigner-Ville Distribution
FIR Finite Impulse Response
FFT Fast Fourier Transform
LPC Linear Predictive Coding
NDE Non Destructive Evaluation
PFA Probability of False Alarm
PD Probability of Detection
PSD Power Spectral Density
ROC Receiver Operating Characteristic
RMS Root Mean Square
SHM Structural Health Monitoring
SNR Signal to Noise Ratio
STFT Short Time Fourier Transform
TSA Time Synchronous Average
vii
LIST OF FIGURES
Figure 1: Angular Resampling Schematic [45]. ........................................................................... 22
Figure 2: Example of Short Delayed Window Structure for use in determining the DRS filter transfer function [51]. ........................................................................................ 24
Figure 3: Residual Difference and Band Pass Mesh Power Spectral Densities for Baseline\Worn tooth Damage @ 1000 RPM Resampled to 2048 Samples per Revolution with 20 revolutions. ........................................................................... 26
Figure 4: CWT scalogram and Hölder exponent series for a baseline condition gearbox signal. ......................................................................................................................... 30
Figure 5: CWT scalogram and Hölder exponent series for a worn tooth condition gearbox signal. ........................................................................................................... 31
Figure 6: Power Spectral Densities of Hölder exponent series for undamaged and worn tooth damaged gearbox signals. ................................................................................. 32
Figure 7: Machinery Fault Simulator (Top View) ........................................................................ 35
Figure 8: Machinery Fault Simulator Controller (Left Side) ........................................................ 35
Figure 9: Gear Box Belt Drive (Side View) ................................................................................. 36
Figure 10: Machinery Fault Simulator Gear Box (Top View) ..................................................... 36
Figure 11: Data Acquisition Card and Sensor .............................................................................. 38
Figure 12: Tachometer Sensor Location (Channel 0) ................................................................... 39
Figure 13: Gear Box Sensor Location (Channel 1) ...................................................................... 40
Figure 14: Bearing Sensor Location (Channels 2-3) .................................................................... 40
Figure 15: Example of hypothesis for damage feature mean greater than baseline feature mean. .......................................................................................................................... 54
Figure 16: Example of hypothesis for damage feature mean less than baseline feature mean. .......................................................................................................................... 54
Figure 17: Example ROC curves for Gaussian Distributions with increasing separation of means. .................................................................................................................... 56
Figure 18: Hypothetical bi-modal (top) and changing variance (bottom) damage distributions. ............................................................................................................... 57
viii
Figure 19: ROC curves of bi-modal damaged distribution (Feature 1) and changing variance damaged distributions (Feature 2). .............................................................. 58
Figure 20: Trial Power Spectrums for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, Channel 1 ........................................................................... 62
Figure 21: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 1 ................................................. 63
Figure 22: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 1 ................ 64
Figure 23: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, 4 Revolutions, Channel 1 ................................................... 64
Figure 24: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 4 Revolutions, Channel 1 .................. 65
Figure 25: Worn Tooth Damage Power Spectrum @ 1000RPM, 512 Samples per Revolution, Channel 1 ................................................................................................ 66
Figure 26: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 512 Samples per Revolution, 20 Revolutions, Channel 1 ................................................. 66
Figure 27: Damage Feature Statistical Performance for Worn tooth Damage Detection @ 1000 RPM, 512 Samples per Revolution, 20 Revolutions, Channel 1 .................. 67
Figure 28: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 512 Samples per Revolution, 4 Revolutions, Channel 1 ................................................... 68
Figure 29: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 512 Samples per Revolution, 4 Revolutions, Channel 1 .................... 68
Figure 30: Trial Power Spectrums for Worn Tooth Damage @ 2000RPM, 2048 Samples per Revolution, Channel 1 ........................................................................... 70
Figure 31: Damage Feature Distributions for Worn Tooth Gearbox Damage @ 2000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1 ...................... 71
Figure 32: Damage Feature Statistical Performance for Worn Tooth Gear Box Damage Detection @ 2000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1 .................................................................................................................... 71
Figure 33: Trial Power Spectra for Main Shaft Bearing Ball Fault Damage @ 1000RPM, 2048 Samples per Revolution, Channel 3 ................................................ 73
Figure 34: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 ...................... 74
ix
Figure 35: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 .............................................................................................. 74
Figure 36: Trial Power Spectrums for Main Shaft Bearing Ball Fault Damage under High Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3 ............. 75
Figure 37: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage Under High Belt Tension @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 .............................................................................................. 76
Figure 38: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Under High Belt Tension Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 ................................................................ 76
Figure 39: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage @ 1000RPM, 2048 Samples per Revolution, Channel 3 ................................................ 78
Figure 40: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 .................................................................................................................... 79
Figure 41: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 .............................................................................................. 79
Figure 42: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage w/ Varying Fault Position @ 1000RPM, 2048 Samples per Revolution, Channel 3 .................................................................................................................... 81
Figure 43: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage w/ Varying Fault Position @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 ...................................................................... 81
Figure 44: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage Detection w/ Varying Fault Position @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 ................................................. 82
Figure 45: Trial Power Spectrums for Main Shaft Bearing Ball Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3 .................................................................................................................................. 84
Figure 46: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 .............................................................................................. 85
x
Figure 47: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Damage Detection w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3 ................................................................ 85
Figure 48: Trial Power Spectrums for Main Shaft Bearing Inner Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3 .................................................................................................................................. 87
Figure 49: Damage Feature Distributions for Main Shaft Bearing Inner Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 ...................................................................... 87
Figure 50: Damage Feature Statistical Performance for Main Shaft Bearing Inner Race Fault Damage Detection w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 ................................................. 88
Figure 51: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3 .................................................................................................................................. 89
Figure 52: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 ...................................................................... 90
Figure 53: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 ...................................................................... 90
Figure 54: Trial Power Spectrums for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, Channel 1 ........................................... 93
Figure 55: Damage Feature Distributions for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1 .............................................................................................. 94
Figure 56: Damage Feature Statistical Performance for Worn Tooth Damage w/ Varying Torsional Load @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1 .............................................................................................. 94
Figure 57: Trial Power Spectrums for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, Channel 3 ........................................... 95
Figure 58: Damage Feature Distributions for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 .............................................................................................. 96
xi
Figure 59: Damage Feature Statistical Performance for Worn Tooth Damage w/ Varying Torsional Load @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3 .............................................................................................. 97
xii
LIST OF TABLES
Table 1: Traditional Damage Features [4], [12], [22], [23]. ......................................................... 28
Table 2: Gear Box Specifications ................................................................................................. 37
Table 3: PCB 352-C33 Accelerometer Specifications ................................................................. 38
Table 4: NI 9234 DAQ Card Secifications ................................................................................... 39
Table 5: System States for Data Sets ............................................................................................ 42
xiii
ACKNOWLEDGEMENTS
I acknowledge and thank my advisor Michael Todd for all his advice, insight,
and discussion over the past two years that has led to the completion of the work
provided in this thesis. Special thanks to Chuck Farrar and the Los Alamos National
Laboratory as well for providing the funding and fellowship. I would like to thank
Dustin Harvey for his support in the acquisition of the data used in this thesis as well as
the development in the condition based maintenance module for SHM tools. I would
like to thank Zhu Mao for his assistance and conversations on various aspects of
statistical analysis. Additional thanks to Colin Haynes, Scott Oullette, Richard Do, and
Eric Kjolsing for their guidance and knowledge they have shared over my career as a
graduate student researcher at UCSD.
xiv
VITA
2008 A.A. University Studies: Pre-Engineering
2010 Engineer-In-Training Certified
2011 B.S Structural Engineering
2013 Master of Science, University of California, San Diego
xv
ABSTRACT OF THE THESIS
A Performance Comparison of Condition Based Monitoring Damage Features Used in Rotating Machines under Variable Conditions
by
Luke Thomas Robinson
Master of Science in Structural Engineering
University of California, San Diego, 2013
Professor Michael Todd, Chair
Condition based monitoring (CBM) is a subset of structural health monitoring
(SHM) that is focused on monitoring vibration signals generated by rotating machines
in situ and processing the data by various techniques designed to extract damage-
sensitive features as a means of performing damage (e.g., bearing and gear failure)
presence, location, or extent. To date, a wide variety of CBM techniques have been
documented in the literature and are well accepted in the CBM community. The
literature provides current technical means for extracting damage features and in some
cases the trending of features in run-to-failure experiments under constant mechanical
xvi
parameters such as load and rotational speed, but it lacks any statistical analysis on the
effects that varying parameters of the mechanical systems for binary damage states have
on detectability. Specifications on data acquisition and choice of algorithm parameters
used in extracting the damage-sensitive features remain somewhat vague. This thesis
attempts to provide a better global understanding of how variability in the damage
detection problem affects the features as a precursor for future work in pattern
recognition and optimal detection of damage to rotating machines. This thesis compares
the various features under varying conditions and computational parameters in a
statistically rigorous way using receiver operating characteristic curves, which compare
the probability of detection vs. the probability of false alarms as a means to improve
detectability for future use in embedded systems. This thesis also introduces a new
damage feature which demonstrates superior detection performance when compared to
traditional damage features for use in detecting worn tooth gear box damage.
1
Chapter 1
Introduction to Structural Health Monitoring and
Condition Based Monitoring of Rotating Machines
1.1. Introduction
Monitoring the health of a system is a multi-faceted problem but, at its most
fundamental level, it seeks to answer the question of, “Is the structure damaged?” (a
form of binary hypothesis test). Structural health monitoring (SHM) is the study of
developing a monitoring system specific to a structures unique characteristics and seeks
to answer this seemingly basic question. SHM is not isolated to any one field of
industry but instead many different fields of study in the pursuit to provide damage
detection capabilities, diagnostic information, and even prognostic information for any
given structure.
Many different application domains have sought SHM solutions. One example
of value to the civil engineering field would be to provide information on the health of
infrastructure after a natural (e.g., earthquake) or human-caused terrorist attack; the civil
engineering community has considered vibration-based monitoring since the early
1980s [1]. The aerospace industry can benefit from structural health monitoring systems
2
by monitoring the health of various aspects of a air and space substructures from wings
to flight systems to engine components; the aerospace community began to study the
use of vibration-based monitoring during the late 1970s and early 1980s [1].
Mechanical engineering applications stand to benefit from the study of SHM as well,
from monitoring machines that are part of an assembly line to engines and braking
systems in vehicles. Even the electrical engineering field has adopted SHM techniques
in monitoring vibration-based damage applied to hard drives disk readers in order to
detect potential writing errors of data and monitor electrical conductivity paths during
chip fabrication. Amongst the different applications there have been varying degrees of
success in the use of SHM. One of the most successful is the study of vibration-based
monitoring of rotating machines known as condition-based monitoring (CBM), which
has been repeatedly applied with success to a variety of machinery elements [2].
Regardless of the SHM application, however, the overall process adopted to accomplish
SHM goals remains essentially the same, as will be described subsequently.
1.2. Structural Health Monitoring Process
The SHM process can be broken down into a four-step statistical pattern
recognition paradigm that provides the foundation for any general SHM application.
This four step process has been outlined in [1][2][3] and is summarized in the list
below along with some of the major considerations involved with each step.
1. Operational Evaluation Process
a. Define cost justifications for SHM implementation.
b. Define system-specific damage types that are to be detected.
3
c. Define operational and environmental conditions under which the
system is to be operated.
d. Define constraints on the data acquisition system or
implementation process under operational conditions.
2. Data Acquisition and Cleansing Process
a. Define the type of data to be acquired and how often it should be
collected.
b. Define the data to be used in the feature selection process.
c. Define requirements for sensor and hardware involved in data
acquisition.
d. Define sensor placement locations.
e. Define sources of variability and normalization procedures to
reduce this variability.
3. Feature Selection Process
a. Define which features of the data are best for damage detection
(sensitivity, robustness)
b. What are the statistical distributions of the extracted damage
features?
c. What type of data condensation can be utilized?
d. Define sources of variability that may affect specific features.
4. Statistical Model Development Process
a. Is data available from undamaged and damaged system
(supervised learning) or only available from undamaged system
(unsupervised learning)?
4
b. Damage Classification
i. Is the system damaged or undamaged?
ii. Where is the damage located?
iii. What type of damage is present?
iv. What is the extent of the damage?
c. Damage Prognostics
i. What is the remaining useful life of the structure?
ii. What is the probability of a correct diagnosis or
prognosis?
The process outlined above may not always be developed in full for any
particular SHM problem, dependent upon constraints, cost, or specific objective(s)
defined. In general, the development of SHM systems are most often driven by cost
justifications.
1.2.1. Operational Evaluation
In the operational evaluation phase of developing a SHM application the first
consideration will be cost. For development to begin the economic benefits of detecting
damage (and subsequently making the right decisions about maintenance, operation,
etc.) must outweigh the cost of development and deployment. Some such things to
consider are the savings that detecting damage would have on maintenance by reducing
down time and, in extreme cases, the benefit that detecting damage would have on life
safety. When defining the system-specific damage types to be detected, cost again is the
driving factor. Structural elements having little effect on its overall performance or for
which the cost of failure is relatively inexpensive may eliminate some failure modes (or
5
monitoring regions) from consideration, whereas even low-probability damage that may
lead to catastrophic failure and cost might drive those decisions. The consideration of
operational and environmental conditions need to be made such that appropriate
damage-sensitive features and system hardware can be selected and still maintain an
adequate level of performance under the given conditions. Lastly in assessing the
limitations and challenges of the data acquisition, one must consider things like the size
of the structure, what part of the structure is accessible for sensor placement, how will
the data or features will be transmitted through the sensor network, and how will the
sensors be powered. Overall the operational evaluation phase involves planning and
development to solve the problems and costs associated with developing the SHM
system.
1.2.2. Data Acquisition
The data acquisition process involves defining and acquiring the raw data types
needed for the application at hand. Common kinematic data types for many structural
applications include temperature, strain, vibrations, or displacements. The type of data
that is to be acquired subsequently drives sensor selection. Infrared sensors or
thermocouples may be used to detect the systems temperature, electrical resistant strain
gauges or fiber optic strain gauges may be used to monitor strain, various
accelerometers can be used for vibration-based monitoring and proximity probes or
laser displacement sensors would be used for detecting displacements.
Data size, sampling frequency, and duty cycle will define what sort of hardware
is needed. In addition to the period associated with data collection, the sampling rate
will also affect what type of hardware that is selected. In the case of monitoring
vibrations, for example, if the hardware and sensor(s) used are only capable of
6
collecting data at 10 kHz, only frequencies of 5 kHz or lower can be resolved using a
Fourier Transform based on the Nyquist sampling criteria. If it is known that high
frequency content is a feature of interest then DAQ hardware capable of sampling at the
appropriate rate must be used. High sampling rates also may require specialized storage
capability or read/write speed, as such all these factors must be considered in the data
acquisition phase of the SHM paradigm.
Fundamental to sensor selection type is where to place them and how many are
required. The common solution to the former question is to place the sensor(s) as close
to the expected damage areas as possible. This location is certainly not always known a
priori, although many applications—particularly in applications such as engine
monitoring, rotating machinery, etc.—“hot spot” knowledge is known, to some degree.
Sensor placement can be determined from physics based models, experimentation, or
experience, but it is not a consideration in this thesis.
1.2.3. Feature Selection
In the feature extraction process it is desirable to pick features that will give the
best performance for the system specific damage to be determined during the
operational evaluation process. In the case of damage classification, ideally a feature
with value(s) that will change for the specific damage type it was designed to detect but
for no other damage or environmental changes is desired. The statistical distributions of
the damage features are also important for detection performance. It is desirable to
select features that have low variability under the operational and environmental
conditions expected from the system and the best possible distinction or separation of
the expected values between damaged and undamaged states. Another desirable trait of
good damage features is reduced dimensionality, the lower the dimensionality of the
7
damage feature the easier it is to store and monitor reducing hardware costs. If using
multiple sensors, data fusion is a tech unique used that combines the data from multiple
sensors in a way that lead to improved detectability of the damage feature. Another
technique used is known as data compression which reduced the dimensionality of the
data collected into a reduced dimensional form that retains the damage specific
information but excludes the non-damage specific qualities of the data collected.
1.2.4. Statistical Model Development
In the statistical model development process of the SHM paradigm, the
statistical characteristics of the selected damage features can be used in such a way that
allows for the decisions on the state of the structure to be made. In order for decision to
be made about the state of the structure some form of machine learning has to be done
in order to make an informed decision. There are two flavors of machine learning which
are classified as supervised and unsupervised. In the case of supervised learning the
algorithm used for the statistical calcification train on undamaged as well as damaged
cases. The more damage classifications provided the more detailed the classification can
be assuming there is a decent amount of separation in the damage feature distributions.
The two most common forms of supervised learning are known as group classification
and regression analysis. In the case of unsupervised learning only the undamaged state
is known and in this case of machine learning it is much more difficult to perform
further classification of damage other than damaged or undamaged. The methods most
commonly used in unsupervised learning are known as outlier detection and novelty
detection. Where the decision of damaged or undamaged is made based on the damage
feature value exceeding some specified confidence interval determined from the normal
operating conditions.
8
On the most basic level the model should be able to distinguish if there is
damage present or not. If possible and the appropriate features were extracted, further
classifications can be made to decide where and what type of damage is present. The
next level of diagnostic information involves determining the extent of the damage.
Lastly is providing a prognosis of the system which is the holy grail of structural health
monitoring. Being able to provide a prognosis implies being able to provide a remaining
useful life of a structure. With an accurate prognosis the parties responsible for
operating and maintain a structure can make informed decision on when to perform
maintenance plan on the replacement of a structure and establish the appropriate
business model that will maximize the use of a structure and potentially save lives.
Of course none of this matters much if the machine learning and statistical
models developed lead to poor decision being made. As such the level of classification
is determined by the statistical performance characteristics of the damage features
themselves. There are two types of poor decisions that can be made which are known as
false positives and false negatives. For example in the case of SHM the null hypothesis
is often damage is present, a false positive would be a damage feature values that
indicates damage when there is none, where as a false negative would indicate no
damage when there is damage present. Often is the case that the latter is the more costly
decision from a Bayes risk point of view as the consequences for not detecting damage
can lead to catastrophic failure or loss of life. On the other hand for false positive the
risk involves losing operational time during maintenance and inspection which need not
be performed in the first place. This knowledge reiterates the point that the best damage
features are those that can distinguish damage and classify the extent of damage with
the highest probability of detection and the least probability of false alarms.
9
1.3. Condition Based Monitoring of Rotating Machines
Condition based maintenance of rotating machinery is a successful and
commonly-used subdivision of the SHM field. Its techniques have been used, many
industries, ranging across manufacturing, automotive, aerospace, and energy sectors. It
has been used to implement health monitoring systems which are useful for life safety
in the case of monitoring drive trains of helicopters [4], [5], where a failure in flight
could have catastrophic consequences, to being used on an online system connected to
inventory networks where ordering parts has become an automated process [6]. By
predicting when parts are going to fail, it is possible to pre-order only the needed parts
and have them readily available for when maintenance is needed. It also has remote use
applications such as monitoring the generators in a wind turbine farm [7]–[11]. In
addition to the diverse applications, condition based monitoring has shown some of the
best results in terms providing robust detection of damage.
1.3.1. Operational Evaluation
When considering the implementation of a condition based monitoring system
the benefits that implementing the system will provide must outweigh the cost of
implementation. There are a number of documents that show strong evidence that CBM
implementation has a very good chance of saving the user money in term of reducing
unwarranted maintenance, reducing downtime of operating machines, and increasing
reliability in the system’s ability to operate [12]–[16]. As such, the implementation of
condition based monitoring for the most part is dependent upon the availability of the
funding to implement it.
The damage to be detected, operational and environmental conditions, and data
acquisition constraints are all going to be system dependent. Some of the more common
10
damage types to be identified are gear and bearing failures. A large body of literature
exists which discuss the damage features with specific applications. Primarily CBM of
rotating machines involves monitoring gears and bearings of machines.
1.3.2. Data Acquisition
The selection of the data acquisition is also going to be system-specific. There
are a number of different sensing techniques that have been used in CBM in order to
monitor rotating machines without interrupting its operations. Some of the most
common techniques are vibration-based techniques, oil/oil debris optical (or similar)
techniques, and infrared thermography [6].
Infrared thermography has been used to monitor slight changes of temperature
to certain components of a machine. The idea behind it is that when bearings wear down
there will be an increase of friction, which emits heat. As such an increase in heat may
be an indication of damage. Some applications of this technique have been to monitor
train bearings and braking systems [17].
Oil debris analysis is a technique that involves monitoring the oil that lubricates
parts of the machine for any debris that may be present. Gears and bearings over time
begin to spall and chip, leaving particles of debris in the oil that can be collected from a
reservoir without having to interrupt the operation of the machine. Detection of metal
particle in the oil would be a sign of damage in this case.
Vibration analysis of rotating machines is the most popular of the three methods
discussed in this section due to its ability to localize damage within the machine based
on known attributes associated with gear mesh frequencies and bearing frequencies.
Vibration analysis also has the ability to provide real time data on the machines
11
condition at any given moment allowing for it to be implemented as an online
monitoring system. It is this class of techniques that this thesis will consider.
1.3.3. Feature Selection
Feature selection for rotating machines can be difficult, as there is a lot of
environmental and operational variability unique to each machine. Things that need to
be considered include the operational speeds the loading and noise level of the
environment. Depending on these variables different methods are available to refine and
clean the data before extracting the desired damage feature. The majority of the CBM
literature has focused on this step of the SHM paradigm.
Pre-processing the vibration signal before extracting a feature is a very common
step. One of the reasons for this is that many times frequencies of the machine present
in the vibration signal are phase locked to the shaft position and can be enhanced or
separated using a method like angular resampling, time synchronous averaging, or
discrete random separation. It is also possible to remove the phase locked frequencies
using these same methods for detecting damage of components that are not phase
locked such as bearing damage frequencies, which are reported to be quasi periodic due
to slip in the bearing housing [17].
Some of the more traditional methods of pre-processing include filtering out
various bandwidths of an angular resampled signal in order to enhance frequency
content that is expected to change during certain damage types. At times it is desirable
to filter out the gear meshing harmonics using band-pass filtering as is the case in the
residual and difference signal. Other pre-processed signals filter out frequency content
in bands between gear meshing harmonics such as the band-pass mesh signal [18]–[22].
12
Other methods involve decomposing the vibration signal into a 2D time-
frequency domain; the continuous wavelet transform, short time Fourier transform, and
Wigner-Ville distribution are some of the more common time frequency methods that
have been used to study damage present in rotating machines [23]–[26]. Some methods
presented in the literature for use in condensing the dimensionality of time-frequency
decompositions involve taking the summation of the frequency components or
computing the Holder exponent as a function of time [27]–[30].
Other methods have been presented that look at the kurtosis of the frequency
domain or dissect it into different bandwidths and look for the maximum kurtosis in
each bandwidth as a way of determining which frequency bands are responsive to
damage[31]–[34]. For bandwidths on interest, demodulation and signal envelope
analysis is often a common technique used to enhance the detection of bearing fault
frequencies [35].
Though there are many techniques for extracting damage features that have been
shown to work well in the literature, setting the parameters and establishing a damage
feature that is intolerant to changes in environmental and operational conditions can still
be an appreciable challenge at times.
1.3.4. Statistical Model Development
Statistical model development applied to rotating machines is the least mature
section of the four steps in the SHM process; some of the work that is being done
involves making prognosis of useful remaining life of machines [36]–[38]. There are
essentially two types of modeling used which are physics based models and data driven
models. In the physics based model, primarily work has been done in predicting crack
13
growth and spalling. In the case of data driven modeling, trending is predicted based on
data taken in known conditions [17].
1.4. Scope of Thesis
The scope of this thesis is to study and report on the detection capabilities of the
various commonly-used CBM damage feature extraction algorithms that have been
described in the academic literature to date and characterize their detection perfection
using experimental data for binary damage states using receiver operating characteristic
curves. The literature to date has primarily involved an extensive pursuit of new
techniques that are proved in theory or using models but lack real world application or
statistically significant results. There have been some articles that have applied the
algorithms to testbeds, but the method of testing usually only considers one test run, or
a very few (due to cost) run-to-failure tests. By performing multiple test runs as is done
in this thesis, it is possible to provide a better understanding of the variability of the
damage features. Some papers have written about some of the uncertainty that exists in
the extraction of the damage features from real world applications; for example, in one
article, tests were performed on a helicopter gear box in flight and it was discovered that
false alarms would occur during strenuous maneuvers [4]. Another article indicated that
some of the damage features would increase then decrease as a certain level of wear
proceeded on the gears and suggested that multiple features be used in order to classify
damage [18]. Lastly, a considerable amount of work has been done on optimal sensor
placement and multisensory data fusion in the scope of CBM [39], [40] but could
benefit from a more informed decision based on known feature variability under
changing environmental and operational conditions.
14
Due to the large variability in damage features under certain conditions and the
lack of recommendations for DAQ specifications it was decided that an important
contribution to the field would be to pursue a study on what damage features may have
improved performance under specified conditions and why certain damage features may
experience increased variability as known events occur, such as increased torsional load
under a difficult flight maneuver or a belt drive loosening over time. This study will
help develop a firm foundation and provide a better overview of which damage features
might experience false alarms under some conditions but not in others. Features that are
not affected by certain environmental conditions may make them better damage features
than others under specified conditions. In general, this thesis is an attempt to lay out the
ground work for future study in developing optimal detectors for classifying damage as
well as providing a platform to continue off of by systematically exploring the damage
features under controlled conditions.
In addition to the detailed comparative study of traditional damage features, a
new damage feature was developed and labeled FMH. The FMH features uses the
frequency spectrum of the Hölder Exponent as a measure of the periodicity of non-
linearities associated with gear mesh harmonics. FMH was compared with traditional
damage features under variable operational conditions and has shown superior
performance compared to traditional damage features used in practice.
15
Chapter 2
Technical Background
2.1. Introduction
In Chapter 2 the technical background needed to understand the algorithms used
in this study are presented and explained in greater detail than previously referred to in
the previous chapter. Some examples are presented as well which will assist in
demonstrating the use and limitations of all the various digital signal processing
techniques that have been used in this study. Firstly, techniques in digital signal
processing will be covered. These digital processing techniques provide the foundation
for some of the more advanced techniques used in condition based monitoring. Then,
more feature extraction techniques are presented which are more specific to CBM of
rotating machinery. Lastly, a review of traditional damage features as well as the new
damage feature will be introduced.
2.2. Digital Signal Processing Techniques
Various digital signal processing techniques are used in the feature extraction
process for condition based monitoring, and to fully understand the reason behind how
16
the features are acquired and why they may experience reduced detection performance,
it is of value to provide some basic understanding of some of the more important
techniques used throughout this thesis as well provide some examples to verify that the
application of the techniques work. This section is dedicated to reviewing relevant
digital signal processing and related technical background used in vibration-based CBM
applications.
2.2.1. Fourier Transform & Power Spectral Density
The Fourier transform is a central pillar of many digital signal processing
applications. Its primary purpose is to transform a signal from the time domain to the
frequency domain by decomposing it to its sinusoidal basis (phase and magnitude). The
complex discrete Fourier Transform can be defined as:
1
1 ( 1)( ) ( ) exp( 2 ( 1) )
N
n
nX k x n j k
N N
, for 1 k N (2.1)
where n is the sample number, N is the total number of samples, and k is the frequency
bin. Using the magnitude of the complex Fourier transform it is possible to determine
the power of a signal associated with a specified frequency. The advantages of the
Fourier transform became even greater when the development of the Fast Fourier
Transform. The Fast Fourier Transform (FFT) uses what are known as twiddle factors
to reduce the number of operations by half if the sampling length is a factor of two
which can have large computational cost savings for large signal lengths.
There are some tradeoffs to consider when using the Fourier transform in
regards to time- frequency resolution. Longer signals provide better frequency
resolution but poorer time resolution whereas short signals have better time resolution
17
and poorer frequency resolution. For example, say a one second signal is sampled at 10
kHz; this would provide a digital signal with 10,000 samples. Because the FFT
frequency range spans a length of 10 kHz, a frequency resolution of 1 kHz can be
achieved. To get better time resolution than 1 second, a shorter length signal must be
used. A shorter length signal would have less samples but the FFT would span the same
frequency range. This shorter signal would lead to poorer frequency resolution as there
are now fewer samples for the same range of frequencies.
Noisy and non-stationary signals suffer from fluctuations of the frequency
components which can cause smearing in the power spectrum and at times-frequency
content within the FFT difficult to resolve. One way to deal with this is to use spectral
averaging to estimate the power spectral density, one of the most common forms of
spectral averaging is Welch’s method [41]. The PSD using Welch’s Method is defined
as:
2
1
1ˆ(f ) ( )K
n kk
LP A n
K U
, where 1
2
0
1( )
L
j
U W jL
, for k = 1, 2… K (2.2)
In Welch’s method the power spectral density consists of first segmenting the
signal into K number of L length segments with a specified number of over lapping
samples; typically the number of overlapping samples is 50-75% of the segment length.
The segments are then windowed to help prevent spectral leakage due to the
discontinuities at the ends of the signal. The Hamming, Hanning and Kaiser windows
are some of the more common window functions used in PSD computations. After
windowing the segments in the time domain the FFT of all the segments are taken and
their magnitudes averaged together. The PSD average then needs to be normalized by
18
L/U to compensate for loss of power due to the windowing. The overlapping also helps
regain some of the lost power due to windowing.
2.2.2. Time-Frequency Analysis
The purpose of time-time frequency analysis is to decompose a signal into its
frequency components as a function of time. There are many different time frequency
decompositions, but this study focuses on three of some of the more popular methods.
The three time-frequency decompositions to be discussed in this section are the Short
Time Fourier Transform (STFT), Linear Predictive Coding (LPC) spectrogram, and the
Continuous Wavelet Transform spectrogram. Each of the mentioned time-frequency
analysis methods have their own different tradeoffs and some examples will be shown.
The STFT quite simply is the FFT of a moving window of a specified length
that is shorter than the length of the signal. The shorter the window the better the
localization of the frequency components are in time with the tradeoff that the shorter
the window, the lower the frequency resolution that can be achieved. The STFT can be
defined as:
( , ) ( ) ( ) exp( 2 )S f x t w t j ft dt
(2.3)
where x is the signal, w is a windowing function, S if the STFT time-frequency
decomposition of the signal x, f is the frequency, and tau is the time delay of the
window function.
The LPC Spectrogram has been used for a long time in speech processing [42].
The advantaged of the LPC spectrogram is that better spectral resolution can be
achieved for short time windows giving it an advantage over the STFT. Like the STFT
19
it involves first sectioning the signal into a series of time-delayed windows.
Autoregressive coefficients are then computed which can be used to compute the
spectrum of the window.
The Continuous Wavelet Transform (CWT) scalogram is also another
alternative to STFT for use in time frequency analysis. The CWT scalogram is different
from the STFT and LPC scalogram due to the fact that time-frequency resolution
tradeoffs change throughout its domain. In the high frequency registers it has very poor
resolution but good time resolution, whereas in the lower frequency range it has good
frequency resolution but poor time resolution. The CWT scalogram can be defined as:
*1,s
tx t dt
ss
, where
2
22 21/42
1c b
t
if t f
b
t e ef
(2.4)
is the mother wavelet, s is the scale factor, and τ is the time lag. The CWT scalogram
can be calculated digitally by developing a filter bank of mother wavelets with scales
varying logarithmically and convolving a mirrored signal with each mother wavelet for
each scale.
There are many mother wavelets, e.g., Morlet, Paul, Shannon, B-Spline and
Daubechies among others. The choice of a mother wavelet dictates time resolution and
frequency resolution in the scalogram. Paul wavelets tend to have good time resolution
and poor frequency resolution whereas B-splines have good frequency resolution but
poor time resolution [43]. Morlet wavelets were used in this paper as they have shown
positive results in detecting transients in signals [28], [29], are easy to implement, and
have a balanced tradeoff between time and frequency resolution.
20
The scale of the wavelet is dictated by the central frequency fc and is varied on a
log scale from a desired minimum frequency to a desired maximum frequency. By
varying of the scales in such a manner the CWT scalogram has a dyadic sampling grid
with increased time resolution at high frequencies and reduced time resolution at low
frequencies.
2.2.3. Signal Envelope and the Hilbert Transform
The Hilbert transform is a way of transforming a signal from its real components
to its analytical form. There are a number of ways to compute the Hilbert transform, but
one of the more cost effective methods is to calculate the FFT of the signal, double its
positive frequency components and remove its negative frequency components. After
performing these steps, the inverse FFT of the modified spectrum will result in the
analytic version of the original signal. The analytic signal can then be used to compute
the enveloped signal by taking the magnitude of the analytic signal. The enveloped
signal can be defined as:
2 2(t) Re[A( )] Im[A( )]E t t (2.5)
where A(t) is the analytic signal.
2.3. Condition Based Monitoring Feature Extraction
Part of the feature extraction process involves manipulating the raw signal in a
way that will later improve the detection of damage and in some cases improve
classification of the damage as well. In Section 2.3 some of the more relevant
21
techniques to condition based monitoring of rotating machines are presented and their
applications discussed.
2.3.1. Angular Resampling
Angular resampling is the process of resampling a vibration signal that was
sampled in equal intervals in the time domain to a signal with equal angular interval
relative to some shaft position in a rotating machine. The purpose of angular resampling
is that often times rotating machines will operate with varying speeds, and during the
feature extraction process it is desirable to track certain frequency components that can
be associated with specific shaft, bearing, and gear frequencies. By using this method it
allows for localization of damaged components at varying speeds. Furthermore,
periodic frequency components at varying shaft speeds will tend to be smeared in the
time domain but remain periodic in the angular domain.
Some systems have shaft encoders and are able to sample at equal angular shaft
interval directly from the DAQ, but often times this is very costly [44]. A less expensive
alternative uses a tachometer, which involves an optical sensor to detect the passing of a
reflector on the shaft, which will generate a pulse for every full rotation of a machines
shaft. Yet another method involves using the vibration signal itself if a known periodic
component is present in the signal and can be correlated with the rotation of the shaft.
The procedure for angular resampling using a tachometer signal is summarized
in the schematic shown in Figure 1. It involves using the tachometer signal to determine
the period crossings of the shaft and using those to crossing to estimate the shaft
position for the remaining samples using some sort of interpolation method. The
original signal is then up sampled by a specified constant L and low passed filtered to
remove any frequency content that may not have been present in the original signal. The
22
value of L is dependent on the desired samples per revolution and the minimum angular
resolution that was present in the tachometer signal, which can be determined from the
estimated angular intervals. After interpolating the time series signal to the angular
domain, the final step is to down-sample the signal to the highest order of interest by
first low pass filtering to prevent aliasing of higher frequency content, and then
decimating by the desired value.
Figure 1: Angular Resampling Schematic [45].
If no tachometer signal is available it is possible to use a periodic gear mesh
frequency component within the signal itself for angular resampling [44]. The process is
similar to the tachometer process but instead of using a tachometer signal to estimate
the shaft position corresponding to each sample of the time domain signal, the shaft
23
phase has to be estimated using prior knowledge of the system being monitored. If the
general rotational frequency of the shaft, the number of teeth present on the gear and the
harmonic order present in the vibration signal are known, that frequency can be band
pass filtered out of the raw signal. Once this has been done, it is possible to compute its
analytical form using a Hilbert transform. From the analytic signal, the phase can be
determined and the shaft position estimated from the phase of a gear mesh harmonic.
Using this estimate of the shaft position, the process is the same as was described for
angular resampling using a tachometer signal.
2.3.2. Time Synchronous Averaging
The name time synchronous averaging (TSA) in context of CBM is a bit of a
misnomer in that the method is not averaging the time domain at all. It is absolutely
crucial to first resample to the angular domain, otherwise it will have little benefit. The
purpose of the TSA technique is to further remove any non-periodic components from a
signal. The TSA signal involves averaging each revolution of the resampled signal
together. The TSA signal can be defined as:
1
0
1(t) ( )
N
TSAn
x x t nTN
(2.6)
Here N is the number of revolutions present in the resampled signal, T is the
number of samples per revolutions and t is representative of the angular position. By
averaging the resampled signal revolutions together, the non-periodic components will
phase cancel with each other, and only the periodic components are left. These periodic
components can be said to be unique to gear specific damage whose teeth mesh
frequencies are strongly correlated to the position of the shaft. The TSA signal can also
24
be used to remove periodic components from an angular resampled signal. This
technique can be achieved by removing the TSA signal from each revolution in the
angular resampled signal by subtraction. Though this method is simple and has been
used and examined for many years in the CBM community [19], [46]–[49], it is also
one of the most erroneous methods for separating discrete and random components
from an angular resampled signal.
2.3.3. Discrete Random Separation
Discrete Random Separation (DRS) is an alternative to self-adaptive noise
cancelation signal processing to filter out periodic components from a vibration signal
[50], [51]. It can be used to separate or remove periodic components from a signal for
feature extraction. The method of performing DRS involves computing the average
cross-power spectrum of a signal by a delayed portion (Figure 2) of itself over the
average auto-power spectrum of the delayed portion.
Figure 2: Example of Short Delayed Window Structure for use in determining the DRS filter transfer function [51].
25
This transfer function can be results in theory with values close to 1 at
frequencies that are periodic in the signal and lesser values for non-periodic
frequencies. The estimate of the optimal filter transfer function can be defined as:
*
, ,1*
, ,1
X ( ) X ( )( )
X ( ) X ( )
K dk M k Mk
K d dk M k Mk
f fH f
f f
(2.7)
Once the estimate of the optimal filter transfer function has been
computed the filter’s impulse response can be computed by taking the inverse FFT and
convolved with the original signal in order to filter out the periodic components. To get
the random component of the vibration signal, the periodic component may be removed
by subtraction, but in order to do this it is important to note that the filter consists of a
phase delay that must be compensated for before removing the periodic component.
2.3.4. The Difference, Residual and Band-pass Mesh Signal
The difference, residual, and band pass mesh signals are all vibration signals
which have undergone some sort of filtering process and are traditionally used in
extracting damage features for condition monitoring of rotating machines [4], [52]–[54].
Computing the residual signal involves filtering out the gear main shaft frequencies and
the gear mesh frequencies but retaining any remaining side bands that may be present in
the signal. For computing the difference signal, again it involves filtering out the main
shaft and a gear mesh frequencies, but in addition to these, the first order side bands are
also to be filtered out. Essentially the bandwidth of the band stop filter becomes wider
to incorporate the first-order frequencies. In the case of processing the band pass mesh
signal, is precisely to band-pass filter the gear mesh frequencies. An example of the
26
power spectral densities for the difference, residual and band-pass mesh signal is
presented in Figure 3. For each plot the left and right halves are representative or a
baseline and worn tooth system state.
Figure 3: Residual Difference and Band Pass Mesh Power Spectral Densities for Baseline\Worn tooth Damage @ 1000 RPM Resampled to 2048 Samples per Revolution with 20 revolutions.
2.3.5. Traditional Damage Features
Three of the classical damage features operate on the conditioned raw signal.
The RMS, is effective in tracking overall noise levels and damage, but is not very good
in determining which component is failing. The crest factor is the ratio of the peak level
Instances
Fre
quen
cy
Power Spectral Density: Residual SignalBaseline and Worntooth Damage
1002003004005000
275481
108135162189216243270297324351378405432459486513540567594621648675702729756783810837864891918945972999
-75
-70
-65
-60
-55
-50
-45
-40
-35
-30
-25
Instances
Fre
quen
cy
Power Spectral Density: Difference SignalBaseline and Worntooth Damage
1002003004005000
275481
108135162189216243270297324351378405432459486513540567594621648675702729756783810837864891918945972999
-75
-70
-65
-60
-55
-50
-45
-40
-35
-30
-25
Instances
Fre
quen
cy
Power Spectral Density: BandPass Mesh SigBaseline and Worntooth Damage
1002003004005000
275481
108135162189216243270297324351378405432459486513540567594621648675702729756783810837864891918945972999
-75
-70
-65
-60
-55
-50
-45
-40
-35
-30
-25
27
of the signal and the RMS. This feature is more sensitive to damage incurred in early
stages of gear and bearing damage. The kurtosis is the fourth statistical moment and a
measure of the peaked-ness of the raw conditioned signals distribution. It is used as an
indicator of major peaks that may appear as damage occurs [54].
FM0 monitors the amplitude of the frequency bins that correspond to the gear
mesh harmonics in the frequency domain of the TSA signal. The PPA is defined as the
peak-to-peak amplitude of the TSA signal. The FM0 is not a good detector for minor
tooth damage [54].
NA4* is an enhanced version of NA4 developed to detect progressive damage.
It normalizes the 4th statistical moment of the residual signal by the variance of a
baseline (healthy) residual signal variance [20], [54].
FM4, M6A, and M8A all operate on the difference signal. FM4 was developed
to detect damage on a limited number of gear teeth. FM4 is 4th statistical moment of the
difference signal normalized by its variance squared [54]. M6A and M8A were
developed to detect surface damage on machinery components. M6A and M8A are
expected to be more sensitive than FM4 [54].
NB4* is computed using the enveloped signal of the band-pass meshed signal. It
is used to detect transients that differ from normal tooth load fluctuations that are
detected by the enveloped signal. The enveloped signal (E) is determined by taking the
magnitude of an analytic signal determined using a Hilbert transform [54] and is
normalized by the variance of a baseline (healthy) enveloped signal.
28
Table 1: Traditional Damage Features [4], [52]–[54].
Damage Feature Feature Equation Signal Used Filtering
Operation
RMS 2
1
1[x ]
N
iiN
Raw Conditioned Signal
--
Crest Factor PeakLevel
RMS Raw Conditioned
Signal --
Kurtosis
4x
12 2
1[x ]
[ ]
N
ii
x
N
Raw Conditioned Signal
--
FM0
1
( )N
ii
PPA
A f
Time Synchronous Averaged Signal
--
NA4*
1N
[ri
r]4
i1
N
( M
2)2
Residual Signal
Band-Stop: Gear Mesh and Shaft
Frequencies
FM4
4
12
2
1
[d ]
[d ]
N
i di
N
i di
N
Difference Signal
Band-Stop: Gear Mesh +1st Order
Sideband and Shaft Frequencies
M6A
2 6i d
13
2i d
1
[d ]
[d ]
N
i
N
i
N
Difference Signal
Band-Stop: Gear Mesh +1st Order
Sideband and Shaft Frequencies
M8A
3 8i d
14
2i d
1
[d ]
[d ]
N
i
N
i
N
Difference Signal
Band-Stop: Gear Mesh +1st Order
Sideband and Shaft Frequencies
NB4*
1N
[Ei
E]4
i1
N
( M
2,E)2
Enveloped Band-Passed Mesh
Signal
Band-Pass: Gear Mesh Frequencies
29
2.3.6. Hölder Exponent and the FMH Feature
The Hölder exponent is computed by calculating the slope of the log-log
analysis of the frequency vs. scale values of the CWT scalogram for each instance of
time. The Hölder exponent as a function of time is defined as:
1
2
mH
, where
log
log
sm
s
(2.8)
Non-linearities caused by tooth impacts of gears generate high frequency energy
in the system. Detecting this increase in high frequency energy can be utilized using the
CWT scalogram and its Hölder exponent, which is known for its sensitivity to local (in
time) non-linearities in dynamic response signals [4,6], making it an ideal candidate for
health monitoring involving the presence or lack of non-linearities in rotating machines.
Healthy gear mesh signals, which consist of periodic tooth contact shocks
dependent on shaft speed, create non-linearities picked up by the Hölder exponent. The
result is a periodic signal associated with the meshing of the gear teeth. By monitoring
the frequency bins of the Hölder exponent power spectral density associated with the
gear mesh harmonics and its peak-to-peak amplitude, it is possible to achieve higher
sensitivity than the traditional methods. The FMH is defined as:
1
( )
HolderN
Holder ii
PPAFMH
A f
(2.9)
An example of the new FMH damage feature for use in detecting worn tooth
damage is provided below. The data were taken from a gearbox mounted sensor with
30
the gear box drive shaft speed operating at approximately 4.49 Hz and 27 teeth per gear.
A single instance of a baseline and damaged CWT scalogram and their corresponding
Hölder exponent series are shown below (Figure 4 and Figure 5).
When computing the CWT scalogram for the healthy gearbox, the gear mesh is
clearly visible in the high frequency bins as can be seen in Figure 4. The corresponding
Hölder exponent series appears to be periodic in nature and similar to an amplitude-
modulated sinusoid with the frequency at the fundamental mesh frequency. For every
tooth mesh impact there is a rise in the high frequency energy content due to the
broadband response of the gear mesh impacts. The high frequency mesh-pattern shows
an expected 27 impacts per full rotation, which agrees with the maximum number of 27
teeth in the drive shaft gear.
Figure 4: CWT scalogram and Hölder exponent series for a baseline condition gearbox signal.
Shaft Rotation (Degrees)
Fre
que
ncy
(Hz)
Morlet CWT Scalogram
0 50 100 150 200 250 300 35045
115297767
19825120
-9 -8 -7 -6 -5 -4 -3 -2 -1 0
0 50 100 150 200 250 300 350-1
0
1Holder Exponent Series
Shaft Rotation (Degrees)
Am
plit
ud
e
31
In the case of the damaged gearbox, the pattern associated with the gear mesh
has clearly decayed (Figure 5). There is now a smearing in the high frequency energy
content instead of the clearly periodic pattern that was seen in Figure 4. When looking
at the Hölder exponent series, the peaks have also been substantially reduced and no
longer resemble the amplitude-modulated sinusoid that was seen in the case of the
healthy gearbox; a decrease in the peak to peak amplitude is also observed.
Figure 5: CWT scalogram and Hölder exponent series for a worn tooth condition gearbox signal.
The power spectral densities of the Hölder exponent series were taken using
Welch’s method. A comparison of the healthy Hölder exponent power spectral density
for one instance of a healthy gearbox and a damaged gearbox can be seen below in
Figure 6. In the case of the healthy gearbox, the first three harmonics can be clearly
seen using a dB amplitude scale. In the case of the damaged gearbox, the amplitudes
drop as the intensity of the gear mesh has reduced due to the tooth wear. The decrease
Shaft Rotation (Degrees)
Fre
qu
en
cy (
Hz)
Morlet CWT Scalogram
0 50 100 150 200 250 300 35045
115297767
19825120
-9 -8 -7 -6 -5 -4 -3 -2 -1 0
0 50 100 150 200 250 300 350-1
0
1Holder Exponent Series
Shaft Rotation (Degrees)
Am
plit
ud
e
32
seen here in the amplitude of frequency bins associated with the gear mesh harmonics
make them a good candidate for a damage detection feature of gearboxes.
Figure 6: Power Spectral Densities of Hölder exponent series for undamaged and worn tooth damaged gearbox signals.
From the example it can clearly be seen that the FMH feature has potential to be
a very effective damage detector for worn tooth damage. The performance of FMH
damage feature is compared to traditional damage features in the results section of this
thesis. Using a statistically significant comparison the FMH feature will be shown to
have superior detection capabilities over traditional methods for early onset gear
damage.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-6
-4
-2
0
2
Holder Exponent PSD
Normalized Frequency (Cycles/Sec)
Am
plit
ud
e (d
B)
HEALTHYWORNTOOTH
33
Chapter 3
Experimental Setup & Feature Extraction Results
3.1. Damage Types and Descriptions
In this study there were four types of damage that were rigorously compared
using the damage features described in the previous chapter. The four types of rotating
machinery damage examined were worn tooth gear box damage, ball fault bearing
damage, outer race bearing damage, and inner race bearing damage. Varying
operational loads such as the tension of the drive belt and torsional gear box load were
also examined. These studies were performed to examine the robustness of the damage
features but were not considered to be damage.
3.2. Experimental Setup
3.2.1. Testbed Description and Evaluation
The data in this study were acquired via the SpectraQuest Inc., Magnum
Machinery Fault Simulator. The fault simulator simulates many faults but the damage
34
types of primary interest for this study were worn tooth (WT) damage on the gearbox
and various bearing faults that were on the main shaft bearing, including ball bearing
outer race faults (BBO), ball bearing inner race faults (BBI), and ball bearing ball faults
(BBB). Some operational parameters that were varied over a number of different tests
were the torsional load (TOR) on the gear box, main shaft speed in revolutions per
minute (RPM), and belt tension (BT) of the belt connecting the gearbox drive shaft to
the main drive shaft. Lastly two different main shaft bearings were used in this study.
System states 1-4 involved using fluid film oil bearing and all other states involved
using ball bearings to support the main shaft.
The testbed consisted of a main shaft driven by an electric motor (Figure 8),
which had a single rotation tachometer signal that indicates the main shaft rotation in
these experiments. The main shaft was set up to be supported by two fluid film oil
bearings for states 1-4 but used ball bearings for the main shaft support for all other
states. The main shaft was a 3/4 inch diameter steel shaft with a support length of 28.5
inches center to center between bearings. Two aluminum masses were supported on the
main shaft, one 8.5 inches to the right from the motor side bearing and the other 6.5
inches from the belt drive side bearing. A belt drive (Figure 9) was used to drive a
gearbox (specifications provided in Table 2) with a magnetic break which applied a
torsional load (Figure 10). The belt drive consisted of two double groove belt sheaves
with a sheave ratio of ~1:3.71 where the smaller sheave was attached to the main shaft
and the larger sheave to the gearbox shaft.
35
Figure 7: Machinery Fault Simulator (Top View)
Figure 8: Machinery Fault Simulator Controller (Left Side)
36
Figure 9: Gear Box Belt Drive (Side View)
Figure 10: Machinery Fault Simulator Gear Box (Top View)
37
Table 2: Gear Box Specifications
Ratio: 1.5:1
Model: Hub City M2
Pitch Angle Gear: 56 degrees 19 minutes
Pitch Angle Pinion: 33 degrees 41 minutes
Pressure Angle for Gear and Pinion: 20 degrees
Material: Forged Steel
Backlash Tolerance: 0.001‐0.005 inches
Pitch Diameter Pinion: 1.125 inches
Pitch Diameter Gear: 1.6875 inches
Number of teeth (Pinion): 18
Number of Teeth Gear: 27
Pinion Bearing: NSK 6202 (1 Bearing)
Gear Bearing: (2 Bearings)
3.2.2. Data Acquisition
The choice in sensors and data acquisition hardware was determined based on a
couple of factors. The first factor was the limitations of the fault simulator. The
literature that came with the hardware recommended that the fault simulator not be run
at speed higher than 6000 RPM for periods longer than 30 seconds as well it was
recommended not to run the near 20% of critical speeds for more than 20 seconds. As
such it was decided so as not to damage any components uncontrollably and run only
1000 RPM and 2000 RPM tests. This limitation meant mesh frequencies and bearing
fault frequencies would be relatively low compared to what the maximum sampling
rates offered by the DAQ systems available in the lab. Lastly, to minimize data storage
space and processing times, lower sampling rates were used.
The sensors chosen were PCB 352-C33 accelerometers, and the DAQ card was
an NI 9234 DAQ Card (Figure 11). Sensor specifications are summarized below in
Table 3 and DAQ card specifications summarized in Table 4. The maximum sampling
38
rate using internal timers for the NI 9234 was 51 kHz whereas the maximum frequency
range was only 0.5-10 kHz for the sensor. Due to the limitations of the sensor all data
were sampled at 10.2 kHz. BNC cables were used in connecting the sensors to the NI
DAQ card, for all data in this study the tachometer (Figure 12) was connected to
Channel 0, the sensor mounted to the gear box (Figure 13) was connected to Channel 1,
the top mounted bearing sensor was connected to Channel 2 and the side mounted
bearing sensor was connected to Channel 3 (Figure 14).
Figure 11: Data Acquisition Card and Sensor
Table 3: PCB 352-C33 Accelerometer Specifications
Sensitivity (±10 %) 100 mV/g
Measurement Range ±50 g pk
Frequency Range (±5 %) 0.5 to 10000 Hz
Frequency Range (±10 %) 0.3 to 15000 Hz
Resonant Frequency ≥50 kHz
Broadband Resolution (1) 0.00015 g rms
Non‐Linearity ≤1 %
Transverse Sensitivity ≤5 %
39
Table 4: NI 9234 DAQ Card Specifications
Number of Channels 4
ADC Resolution 24 bit
Input Range ±5 V
Sampling Mode Simultaneous
Data Range Min (Internal Master Time Base) 1.652 kS/s
Data Range Max (Internal Master Time Base) 51.2 kS/s
Figure 12: Tachometer Sensor Location (Channel 0)
40
Figure 13: Gear Box Sensor Location (Channel 1)
Figure 14: Bearing Sensor Location (Channels 2-3)
41
After connecting the sensors to the NI card, a USB cable was used to connect
the NI card was connected computer. A MATLAB program developed by a colleague
Dustin Harvey was used to collect data from the testbed during operation.
3.2.3. Description of System States
During the course of this study a number of data sets were collected over several
days in some cases and others were taken in 1 hour sessions depending on what purpose
each set served. Two procedures were used for the collection of data. The first
procedure used for collecting the preliminary data sets (1-9) was not as formalized in an
attempt to let environmental variability occur in the process and observe how that
variability manifested itself in the data. The second procedure used was a more
formalized procedure in an attempt to reduce environmental variability associated with
assembly and breakdown. The various states collected for use in this study are
summarized in Table 5.
42
Table 5: System States for Data Sets
State Number
Main Shaft Bearing
Damage Indicator
Damage Type
Main Shaft Speed (RPM)
Belt Tension (BT)
Break Torsion (TOR)
Number of Trials
1 Fluid 0 BASE 1 ~MID 1 256
2 Fluid 1 WT 1 ~MID 1 256
3 Fluid 0 BASE 2 ~MID 1 256
4 Fluid 1 WT 2 ~MID 1 256
5 Ball 0 BASE 1 ~MID 1 256
6 Ball 1 BBB 1 ~MID 1 256
7 Ball 1 BBB 1 ~HIGH 1 192
8 Ball 1 BBO 1 ~MID 1 256
9 Ball 1 BBOV 1 ~MID 1 256
10 Ball 0 BASE 1 2 1 128
11 Ball 0 BASE 1 4 1 128
12 Ball 0 BASE 1 4 3 128
13 Ball 0 BASE 1 4 5 128
14 Ball 0 BASE 2 4 1 256
15 Ball 1 BBB 1 2 1 128
16 Ball 1 BBB 1 4 1 128
17 Ball 1 BBI 1 2 1 128
18 Ball 1 BBI 1 4 1 128
19 Ball 1 BBO 1 2 1 128
20 Ball 1 BBO 1 4 1 128
21 Ball 1 WT 1 4 1 128
22 Ball 1 WT 1 4 3 128
23 Ball 1 WT 1 4 5 128
State 1 consisted of 256 trials and was a baseline condition to be used for
supervised learning. The main shaft was supported by fluid film bearings and its speed
was monitored using a tachometer. The main shaft speed was approximately 1000 RPM
or 16.7 Hz. The belt tension was set to an approximated medium tension level and the
magnetic brake was set to apply 2lb-in of tension to the undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
43
State 2 consisted of 256 trials and was a damage gear box condition to be used
for supervised learning. The main shaft was supported by fluid film bearings and its
speed was monitored using a tachometer. The main shaft speed was approximately 1000
RPM or 16.7 Hz. The belt tension was set to an approximated medium tension level and
the magnetic brake was set to apply 2lb-in of tension to a gear box pinion. The gear box
here suffered from worn teeth. Testing procedure 1 was used for the data collection.
State 3 consisted of 256 trials and was a baseline condition to be used for
supervised learning. The main shaft was supported by fluid film bearings and its speed
was monitored using a tachometer. The main shaft speed was approximately 2000 RPM
or 33.3 Hz. The belt tension was set to an approximated medium tension level and the
magnetic brake was set to apply 2lb-in of tension to the undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
State 4 consisted of 256 trials and was a damage gear box condition to be used
for supervised learning. The main shaft was supported by fluid film bearings and its
speed was monitored using a tachometer. The main shaft speed was approximately 1000
RPM or 16.7 Hz. The belt tension was set to an approximated medium tension level and
the magnetic brake was set to apply 2lb-in of tension to a gear box pinion. The gear box
here suffered from worn teeth. Testing procedure 1 was used for the data collection.
State 5 consisted of 256 trials and was a baseline condition to be used for
supervised learning. The main shaft was supported by ball bearings and its speed was
monitored using a tachometer. The main shaft speed was approximately 1000 RPM or
16.7 Hz. The belt tension was set to an approximated medium tension level and the
magnetic brake was set to apply 2lb-in of tension to an undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
44
State 6 consisted of 256 trials and was a damaged main shaft ball bearing
condition to be used for supervised learning. The main shaft was supported by a
damaged ball bearing with a ball fault and its speed was monitored using a tachometer.
The main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was
set to an approximated medium tension level and the magnetic brake was set to apply
2lb-in of tension to an undamaged gear box pinion. Testing procedure 1 was used for
the data collection.
State 7 consisted of 192 trials and was a damaged main shaft ball bearing
condition to be used to examine the effects of an overly tensioned belt drive. The main
shaft was supported by a damaged ball bearing with a seeded ball fault and its speed
was monitored using a tachometer. The main shaft speed was approximately 1000 RPM
or 16.7 Hz. The belt tension was set to an approximated high tension level and the
magnetic brake was set to apply 2lb-in of tension to an undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
State 8 consisted of 256 trials and was a damaged main shaft ball bearing
condition to be used for supervised learning. The main shaft was supported by a
damaged ball bearing with a seeded outer race fault and its speed was monitored using a
tachometer. The outer race fault position in this case was constant and located closest to
the Channel 3 bearing sensor. The main shaft speed was approximately 1000 RPM or
16.7 Hz. The belt tension was set to an approximated medium tension level and the
magnetic brake was set to apply 2lb-in of tension to an undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
State 9 consisted of 256 trials and was a damaged main shaft ball bearing
condition to be used to examine the effect of varying position of an outer race fault in
45
relation to the sensor and load. The main shaft was supported by a damaged ball bearing
with a seeded outer race fault and its speed was monitored using a tachometer. After
every 32 trials the bearing was disassembled and reassembled with not particular
attention being made as to the position of the where the fault was located in relation to
the sensors or the loading of the shaft. The main shaft speed was approximately 1000
RPM or 16.7 Hz. The belt tension was set to an approximated medium tension level and
the magnetic brake was set to apply 2lb-in of tension to an undamaged gear box pinion.
Testing procedure 1 was used for the data collection.
States 1-9 were collected using testing procedure 1 and after examination of the
performance of the damage features, it was notices that there was a considerable amount
of variability in the damage feature statistical distributions. As such a new testing
procedure was developed to reduce variability that may be cause due to maintenance
issues such as the dismantling of the various damaged components and reassembly as
was done in the preliminary results. By reducing this variability it was hoped that the
distributions of the damage features would become less spread out and the variability
due to torsional load and belt tension could be singled out. States 10-24 used the second
testing procedure described in 3.2.4Testing Procedures.
State 10 consisted of 128 trials and was a baseline condition to be used for
supervised learning of low tensioned belt drive. The main shaft was supported by ball
bearings and its speed was monitored using a tachometer. The main shaft speed was
approximately 1000 RPM or 16.7 Hz. The belt tension was set using a deflection force
of 2 lbs. for 13/64” of an inch deflection and the magnetic brake was set to apply 2lb-in
of tension to the undamaged gear box pinion. Testing procedure 2 was used for the data
collection.
46
State 11 consisted of 128 trials and was a baseline condition to be used for
supervised learning of low torsional load on the gear box. The main shaft was supported
by ball bearings and its speed was monitored using a tachometer. The main shaft speed
was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the recommended
value using a deflection force of 4 lbs. for 13/64” of an inch deflection and the magnetic
brake was set to apply 2lb-in of tension to the undamaged gear box pinion. Testing
procedure 2 was used for the data collection.
State 12 consisted of 128 trials and was a baseline condition to be used for
supervised learning of medium torsional load on the gear box. The main shaft was
supported by ball bearings and its speed was monitored using a tachometer. The main
shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the
recommended value using a deflection force of 4 lbs. for 13/64” of an inch deflection
and the magnetic brake was set to apply 6lb-in of tension to the undamaged gear box
pinion. Testing procedure 2 was used for the data collection.
State 13 consisted of 128 trials and was a baseline condition to be used for
supervised learning of high torsional load on the gear box. The main shaft was
supported by ball bearings and its speed was monitored using a tachometer. The main
shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the
recommended value using a deflection force of 4 lbs. for 13/64” of an inch deflection
and the magnetic brake was set to apply 10lb-in of tension to the undamaged gear box
pinion. Testing procedure 2 was used for the data collection.
State 14 consisted of 256 trials and was a baseline condition to be used for
supervised learning of low torsional load on the gear box at higher speeds. The main
shaft was supported by ball bearings and its speed was monitored using a tachometer.
47
The main shaft speed was approximately 2000 RPM or 33.3 Hz. The belt tension was
set to the recommended value using a deflection force of 4 lbs. for 13/64” of an inch
deflection and the magnetic brake was set to apply 2lb-in of tension to the undamaged
gear box pinion. Testing procedure 2 was used for the data collection.
State 15 consisted of 128 trials and was a damaged condition to be used for
supervised learning ball fault bearing failure at low belt tension. The main shaft was
supported by ball bearings and its speed was monitored using a tachometer. The main
shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the
low setting value using a deflection force of 2 lbs. for 13/64” of an inch deflection and
the magnetic brake was set to apply 2lb-in of tension to the undamaged gear box pinion.
Testing procedure 2 was used for the data collection.
State 16 consisted of 128 trials and was a damaged condition to be used for
supervised learning ball fault bearing failure at recommended belt tension. The main
shaft was supported by ball bearings and its speed was monitored using a tachometer.
The main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was
set to the recommended setting value using a deflection force of 4 lbs. for 13/64” of an
inch deflection and the magnetic brake was set to apply 2lb-in of tension to the
undamaged gear box pinion. Testing procedure 2 was used for the data collection.
State 17 consisted of 128 trials and was a damaged condition to be used for
supervised learning inner race ball bearing failure at low belt tension. The main shaft
was supported by ball bearings and its speed was monitored using a tachometer. The
main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to
the low setting value using a deflection force of 2 lbs. for 13/64” of an inch deflection
48
and the magnetic brake was set to apply 2lb-in of tension to the undamaged gear box
pinion. Testing procedure 2 was used for the data collection.
State 18 consisted of 128 trials and was a damaged condition to be used for
supervised learning inner race ball bearing failure at recommended belt tension. The
main shaft was supported by ball bearings and its speed was monitored using a
tachometer. The main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt
tension was set to the recommended setting value using a deflection force of 4 lbs. for
13/64” of an inch deflection and the magnetic brake was set to apply 2lb-in of tension to
the undamaged gear box pinion. Testing procedure 2 was used for the data collection.
State 19 consisted of 128 trials and was a damaged condition to be used for
supervised learning outer race ball bearing failure at low belt tension. The main shaft
was supported by ball bearings and its speed was monitored using a tachometer. The
main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to
the low setting value using a deflection force of 2 lbs. for 13/64” of an inch deflection
and the magnetic brake was set to apply 2lb-in of tension to the undamaged gear box
pinion. Testing procedure 2 was used for the data collection.
State 20 consisted of 128 trials and was a damaged condition to be used for
supervised learning outer race ball bearing failure at recommended belt tension. The
main shaft was supported by ball bearings and its speed was monitored using a
tachometer. The main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt
tension was set to the recommended setting value using a deflection force of 4 lbs. for
13/64” of an inch deflection and the magnetic brake was set to apply 2lb-in of tension to
the undamaged gear box pinion. Testing procedure 2 was used for the data collection.
49
State 21 consisted of 128 trials and was a baseline condition to be used for
supervised learning of low torsional load on a damaged gear box. The main shaft was
supported by ball bearings and its speed was monitored using a tachometer. The main
shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the
recommended value using a deflection force of 4 lbs. for 13/64” of an inch deflection
and the magnetic brake was set to apply 2lb-in of tension to the damaged gear box
pinion with worn gears. Testing procedure 2 was used for the data collection.
State 22 consisted of 128 trials and was a baseline condition to be used for
supervised learning of medium torsional load on a damaged gear box. The main shaft
was supported by ball bearings and its speed was monitored using a tachometer. The
main shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to
the recommended value using a deflection force of 4 lbs. for 13/64” of an inch
deflection and the magnetic brake was set to apply 6lb-in of tension to the damaged
gear box pinion with worn gears. Testing procedure 2 was used for the data collection.
State 23 consisted of 128 trials and was a damaged condition to be used for
supervised learning of high torsional load on the damaged gear box. The main shaft was
supported by ball bearings and its speed was monitored using a tachometer. The main
shaft speed was approximately 1000 RPM or 16.7 Hz. The belt tension was set to the
recommended value using a deflection force of 4 lbs. for 13/64” of an inch deflection
and the magnetic brake was set to apply 10lb-in of tension to the undamaged gear box
pinion. Testing procedure 2 was used for the data collection.
3.2.4. Testing Procedures
The two testing procedures used in the collection of data for this study are
discussed in further detail here. In system states 1-9 the testing procedure used was
50
designed to be informal and allow environmental variability into the data. The second
testing procedure was designed to reduce random environmental variability by not
swapping components between tests and quantifying belt tension and torsional load.
Though designed to be informal, basic guidelines were established for the first
procedure. Several parameters were always kept constant during the preliminary data
collection for states 1-9. The first parameter kept constant was the location of the
sensors. The tachometer was always positioned directly above the main shaft and
monitored the main shaft rotational speed. The accelerometers monitoring the vibration
signals were always mounted at the same locations and in the same way. Channel 1 was
always mounted on top of the gear box, Channel 2 was always mounted on the top of
the second main bearing housing closest to the gear box and belt drive, and Channel 3
was always positioned to the side of the main shaft bearing housing closest to the gear
box and belt drive. The accelerometers were always mounted in the same fashion, using
stud mounting with a wax film to provide good contact between the bottom of the
accelerometer and the surface of the testbed, then hand tightened.
During the course of a trial set using testing procedure 1, the machine fault
simulator was allowed to warm up at the desired speed for the experiment for a period
of ten minutes before any data were collected. Data were collected using the seDAQ
program a single set at a time by hand and the time interval between each instance
varied. The sets were always collected in groups of 32 instances and total time of
collection could range anywhere from 30-45 minutes. The trial sets of 32 instances later
were concatenated to produce a larger state set.
In between trial sets there involved a large amount of modification to the fault
simulator. In the case of gear box fault sets, the worn tooth and baseline gear boxes
51
were two separate gearboxes, between each set the belt drive had to be loosened and
belts removed. Then the gearbox was unfastened and removed, there was only one large
sheave so that too had to be removed and swapped to the new gear box for the next set.
When replacing the gear box the belt drive had to be tensioned with an approximated
level of tension using a ruler and estimated feel of the force being applied.
In the case of bearing damage trial sets, the belt drive again had to be loosened
and removed, the bearing housing was dismantled, and bearings removed to be swapped
with a new bearing. Once the new bearing was replaced, the belts were put back in
place and re-tensioned using the same approximated method as was done for gear box
trial sets.
For States 10-24 the amount of variability due to dismantling and moving
around components was undesirable due to the fact that the purpose of the experiments
was to study the effect of varying mechanical states such as torsional load and belt
tension, not variability due to maintenance and repairs. As such the procedure was
changed from testing procedure 1. The program seDAQ was modified to change the
way each instance of the state was collected. Each instance was no longer collected at
varying time intervals but instead was collected at specified times controlled by the
computer’s internal clock.
Between each state set the system was allowed first to run for ten minutes before
the first trial set of 32 instances were collected. Trial set were spaced at 10 minutes
apart and each instance of a trial set was spaced at 5 seconds apart. State sets were
collected in groups of 4 trial sets and during this time the machine never stopped and no
components were altered by swapping parts, changing the torsional load or belt tension.
52
The method for determining belt tension was also changed to reduce any
variability associated with the approximated method using feel and a ruler. Here in
procedure 2 a belt tension tester was used which measured the force applied to the belt
for a given deflection. The manufacturer of the belts provided recommended deflection
to force ratios for specified belts and span lengths to achieve recommended belt
tensions. In the case of low belt tension system states the force was reduced by half for
the same deflection. To change the torsional loads two screws on the magnetic break
needed to be loosened, the break was rotated to a clearly marked setting and the screws
were retightened.
3.2.5. Sources of Variability
There were a number of sources of variability in the system. One major
contributing factor of variability was the substitution of damaged and undamaged
components between the collections of data sets. The substitution of these components
effected belt tension and caused it to change on a case by case basis. Changing the
gearbox and the sheaves also led to slight changes in the way that the belts were lined
up. Procedures were used to insure the belt drive was orientated in the correct manner,
but slight changes from its previous position was unavoidable and suspected to be a
contributing factor to variability in the statistical distributions of the damage features.
Variability due to operational conditions was also observed. The belt tension
tended to loosen during operation as the belt heated up during longer runtimes. The
force applied for a deflection of 13/64 inches was typically a pound less than it was at
the start of a trial for the longer runs. The speed of the main shaft also varied over the
course of a data collection session. Though the motor controller was set to a constant
power level, the RPM of the main shaft fluctuated by roughly 1-2 RPM.
53
3.3. Statistical Modeling and Detection Performance
To better understand the results section a brief over view of how the damage
features performance was compared is explained here. The damage features were
computed using some of the recommended techniques in the literature, but their values
would vary for every measurement of the signal as the environment changed. The
purpose of the various states that were explained earlier was to see how the different
states affected the statistics of the damage features. A good damage feature is one that
detects a specific condition of the system and whose value does not change excessively
under potential changes in the environment except for changes that indicate the damage
it was designed to detect (“statistical specificity”). In order to quantify the ability to
detect this damage receiver operating characteristics (ROC) curves were used.
3.3.1. Hypothesis Testing & Receiver Operating Characteristic Curves
In order to generate an ROC curve first a null hypothesis needs to be made. For
the results presented the hypotheses answered the question of is there damage present.
Two hypotheses were possible as the data were trained on a binary case where a
baseline state was always compared with a known damage states under different
environmental conditions. The hypotheses were determined by an automated algorithm
where the means of the statistics for the baseline data set and the damaged data set.
Then using this knowledge a hypothesis could be made.
When the mean of the damage feature was larger than the mean of the baseline
feature, the hypothesis leading to the best probability of detection was that damage was
present if the feature value was larger than a specified threshold (Figure 15). When the
mean of the baseline damage features was greater than the mean of the damage features,
the hypothesis was made that damage was present if the feature values were less than
54
some specified threshold (Figure 16). Two visualizations of the different hypothesis are
presented in the figures below where the green is representative of a baseline damage
features distribution and the red a damaged features distribution.
Figure 15: Example of hypothesis for damage feature mean greater than baseline feature mean.
Figure 16: Example of hypothesis for damage feature mean less than baseline feature mean.
With a known hypothesis and known distributions for the damage features it is
possible to generate an ROC curve. ROC curves provide information on the probability
of false alarm (PFA) verse the probability of detection (PD) for varying thresholds. The
PFA can be defined as the probability that an undamaged system may generate feature
Damage Present
Damage Present
PFA
PFA
PD
PD
55
values that would indicate damage, or that a damaged system may generate values that
would not indicate damage for a given hypothesis.
In the case of the experiments, the analytical distributions were unknown; as
such no analytical ROC curves could be generated. In order to generate the ROC
curves, an empirical method was used. As mentioned earlier the hypotheses were made
based on the mean values of known damaged and undamaged states under the same
environmental conditions. After a hypothesis was established a threshold was moved
along the range of features values and the number of feature values that generated false
alarms were counted and plotted against the number of feature values which indicated
damage by satisfying the hypothesis made. In this way the performance of the damage
features could be quantified and compared to each other.
When reading an ROC curve plot the best performing damage feature is the one
with the greatest PD to PFA. Features with distributions that have less variance and
greater separation in their means give the best performance. The ROC curves of the
better performing damage features will tend to shift to the upper right corner of the plots
which is representative of higher PD to PFA ratios. When the features tend to be poor
performers due to little distinction in the statistics of the distributions the performance
will be close to a straight line that runs from one corner of the plot to the other. This line
is the 50/50 line whose performance is no better than a coin toss. In Figure 17 ROC
plots were generated for 5 hypothetical features. The features were of Gaussian
distribution for both damaged and undamaged states. The separation in the means
between the damaged and undamaged conditions for the five features was increased by
a factor of 1 for every feature stating at 0. From the ROC curve it can clearly be seen
that features with large separation in statistics between the damaged and undamaged
56
condition move to the upper right of the ROC plot as was the case with Feature 5. For
Feature 1, the distributions were exactly the same and little information is available to
distinguish the system as being damaged or undamaged.
Figure 17: Example ROC curves for Gaussian Distributions with increasing separation of means.
For certain distributions types, the ROC curves generated from using a single
moving threshold may generate abnormal ROC curves. Such distributions include bi-
modal distributions or distributions with the same mean and different variances (Figure
18). In the case of the bi-modal damage state distribution compared to the single mode
distribution, the ROC curve begins to develop an S shape curve. This same ROC curve
shape can be seen in the case of two distributions with the same mean but differing
variances (Figure 19). When establishing hypotheses from these distribution types a
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
Feature 1
Feature 2
Feature 3Feature 4
Feature 5
57
multi threshold approach should be used in the hypothesis test. One approach would be
to hypothesize that damage is present given the feature value is not within a certain
range between two thresholds. This range can then be increased to provide a ROC curve
for that feature.
Figure 18: Hypothetical bi-modal (top) and changing variance (bottom) damage distributions.
-6 -4 -2 0 2 4 60
0.05
0.1
0.15
Feature Value Probability Distribution
Pro
babi
lity
Feature Value (x)
p(x|state = undamaged) p(x|state = damaged)
-6 -4 -2 0 2 4 60
0.05
0.1
0.15
Feature Value Probability Distribution
Pro
babi
lity
Feature Value (x)
p(x|state = undamaged) p(x|state = damaged)
58
Figure 19: ROC curves of bi-modal damaged distribution (Feature 1) and changing variance damaged distributions (Feature 2).
Depending on the cost of a false alarm vs. the cost of detection, thresholds can
be determined and implemented in a CBM system. In the absence of any cost functions
for a particular application, the area under the curve (AUC) of an ROC curve is often
used as a means of evaluating machine learning algorithms using a scalar metric [55].
The AUC is computed by integrating the ROC curve over the total PFA range. Damage
features with better statistical performance will have AUC values closer to 1 whereas
poor performing damage features will have a value closer to 0.5. Damage features with
an AUC value less than 0.5 indicate a poor choice in hypothesis. In the case that the
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
Feature 1
Feature 2
59
AUC is less than 0.5 swapping the hypothesis would regain any statistical advantage in
the damage features ability to detect damage. The calculation of the AUC for empirical
data can be performed by two methods. One method involves fitting a polynomial to the
ROC data and integrating along the polynomial. A more direct method involves using
trapezoidal integration. In the case of this study the trapezoidal method was used for
computing the AUC of all ROC curves.
3.4. Testing Results
A number of tests were performed to determine the effects that varying system
states had on damage feature detection performance. Testing procedures and system
states were mentioned in previous sections and the results of traditional damage features
will be discussed further in this section. Preliminary results are provided which
compared damage feature detection performance under a number of different damage
classes as well as how different system parameter affected their performance. Two
types of damage were examined and included worn tooth damage in the gearbox, ball
fault damage in the main shaft bearing and outer race fault in the main shaft. DAQ
limitation will dictate signal parameters like sampling frequency and signal length. The
performance of worn tooth damage features were examined under varying signal
parameters that consisted of sampling frequency and number of revolutions present in
the angular resampled signal. Operational parameters also had an effect on damage
detection performance for worn tooth damage. The effect to damage detection
performance caused by increased shaft speed for worn tooth damage detection is also
examined. The performance of bearing damage detection was examined for ball fault
bearing damage and outer race bearing damage. Ball fault bearing damage was
60
examined under medium belt tension and high belt tension. Outer race bearing fault
damage detection was examined when the fault location was randomized and unknown
as well as having a constant position relative to the tension load of the belt.
Because the preliminary results showed there was a lot of variability in the
power spectra from trial to trial. Subsequent test were performed that minimized the
variability associated with substituting parts with different damage classifications and
taking data in short sessions on multiple days in an attempt to identify the extent of
variability caused by specified operational states. The effect of belt tension on bearing
damage detection was examined further for three ball bearing faults: Inner Race, Outer
race and Ball Fault. The affect torsional load had on detecting worn tooth damage and
which damage features detection performance at alternate sensor locations on the
structure was examined as well.
For all the results shown below the only parameters changed in the computation
of damage features were the samples per revolution used in the angular resampling
algorithm, the number of revolutions kept in the signal, the channel the data were
collected from, and the data sets used in the computations; these are all labeled in the
figure titles for clarity.
To compute the damage features the data were loaded and the mean values of
the signals were removed giving them a zero DC value at zero frequency. The signals
were then angular resampled at varying rates to see how the sampling rate would affect
the results of the damage detectability. The following examples were not time
synchronous averaged, as it was observed that this had a reduction of performance for
the damage features in earlier experiments. For the power spectrum plots the power
spectrums were computed for all the samples and the average power spectrum is
61
displayed for both the baseline conditions and damaged conditions. Once all features
had been computed there distributions were plotted using box plots as the actual
statistical distribution was unknown. Damaged classes were plotted in red and
undamaged classes were plotted in green. ROC curves were plotted for each damage
feature using empirical results and the area under the curve was calculated to provide a
measure of total probability of detection. The values of the area under the curve are
presented as green and red arrow to provide some knowledge of what hypotheses were
used for the computation of the ROC curves. The determination of the null hypotheses
used was an automated process determined by the mean of the damaged and undamaged
features in relation to each other. Red arrows pointing down represented hypotheses of
the structure was damaged given the feature value was lower than the specified
threshold, whereas green arrows represented hypotheses of the structure was damaged
given the feature value was greater than the threshold.
3.4.1. Worn Tooth Gearbox Damage Preliminary Results
In this section preliminary gear box system states are compared for a baseline
and worn tooth damage cases to see how the power spectrums are affected by worn
tooth gears during operation. It was desirable to see the performance of the damage
features at varying signal lengths and sampling rates. Damage feature comparisons are
also made at different operating speeds.
The frequency ranges of all power spectrum plots are marked by gear mesh
harmonics. A value of one represents the fundamental meshing frequency which is 27
cycles per rotation of the gear shaft as the gear had 27 teeth. The fundamental meshing
frequency is rarely present in all the experiments taken at 1000 RPM but does become
more apparent at speeds of 2000 RPM. From looking at the PSD’s of the damage cases
62
it is clear that the wearing of the gear teeth caused the frequency content to smear into
other frequency bins and the peaks associated with the gear meshing begin to decay.
There is more high frequency content in the undamaged case due to the contact shocks
of the gear teeth meshing together, whereas in the case of the worn teeth more slipping
occurs and there is less force per contact (Figure 20).
Figure 20: Trial Power Spectrums for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, Channel 1
The ROC curves for the damage features at a sampling rate of 2048 samples per
revolution (Figure 22) showed that the dominant performing damage feature was FMH,
followed by FM4 and M6A. These damage features had good separation of the
damaged and undamaged distributions (Figure 21) giving them better detectability. The
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityWorntooth Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
63
number of revolutions used in the calculations of the damage features was reduced to
only 4 revolutions. The damage feature distributions showed an increase in variance of
the damaged and undamaged feature values when 4 revolutions were used (Figure 23).
This increased variance caused a decrease in detection performance for damage features
FM4, M6A and M8A observed from their ROC curves (Figure 24). The FMH damage
feature continued to have good detection performance despite the reduction in signal
length.
Figure 21: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 1
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
4
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
64
Figure 22: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 1
Figure 23: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 2048 Samples per Revolution, 4 Revolutions, Channel 1
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
1
2
3
4
5
6
7
8
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
65
Figure 24: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 4 Revolutions, Channel 1
The effect of reduced sampling rate on worn tooth damage detection
performance was examined by first reducing the sampling rate to 512 samples per
revolution. The PSD’s of the trials (Figure 25) now only consisted up to 9 gear meshing
orders but nothing appeared to change otherwise. The distributions of the damage
features (Figure 26) showed overlap between most of the damaged and undamaged
states. The ROC curves (Figure 27) showed that for a signal length of 20 revolutions
sampled at 512 samples per revolution, most of the damage features suffered a
reduction in performance compared to the features extracted from signals sampled at
2048 samples per revolution. The two damage features that did not suffer a performance
loss were NA4* and FMH. When reducing the sampling frequency, NA4* had a large
increase in performance and the FMH damage feature continued to perform well.
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6AM8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
66
Figure 25: Worn Tooth Damage Power Spectrum @ 1000RPM, 512 Samples per Revolution, Channel 1
Figure 26: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 512 Samples per Revolution, 20 Revolutions, Channel 1
-40 -30 -20 -100
1
2
3
4
5
6
7
8
9
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 100 150 200 2500
1
2
3
4
5
6
7
8
9
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100
1
2
3
4
5
6
7
8
9
Mean Amplitude (dB)InstancesF
requ
ency
Power Spectral DensityWorntooth Condition
50 100 150 200 2500
1
2
3
4
5
6
7
8
9
-40
-35
-30
-25
-20
-15
-10
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
67
Figure 27: Damage Feature Statistical Performance for Worn tooth Damage Detection @ 1000 RPM, 512 Samples per Revolution, 20 Revolutions, Channel 1
Signals sampled at 512 samples per revolution were cropped to 4 revolutions to
see which feature would perform well under short lengths as well as a reduced sampling
frequency. The damage feature distributions (Figure 28) showed an increase in outliers
and an increase in variance for damage features FM4 M6A and M8A. The ROC curves
(Figure 29) showed that FM4, M6A and M8A had a reduction in performance. The
NA4* and FMH feature remained the better performing damage features when
compared to longer signal lengths at the same sampling frequency.
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
68
Figure 28: Damage Feature Distributions for Worn Tooth Damage @ 1000RPM, 512 Samples per Revolution, 4 Revolutions, Channel 1
Figure 29: Damage Feature Statistical Performance for Worn Tooth Damage Detection @ 1000 RPM, 512 Samples per Revolution, 4 Revolutions, Channel 1
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
1
2
3
4
5
6
7
8
9
10
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
69
Out of all the damage features the top performing feature for detecting worn
tooth damage at a main shaft speed of 1000 RPM was the FMH damage feature as it
demonstrated good detection for both long and short signal lengths as well as multiple
sampling frequencies. This feature property is beneficial because fewer false alarms will
occur and the dependability of the damage feature is increased. The increased
dependability for multiple sampling rates and signal lengths greatly reduces limitations
that may arrive when determining which hardware to acquire or sensors to choose.
Data were taken at shaft speeds of 2000 RPM to observe how the damage
features would perform at increased rotational speeds. The raw vibration signals were
resampled to 2048 samples per revolution which required up sampling the original
frequency content by an approximate factor of 2. The signal length in these trials
consisted of 40 revolutions as a result of the increased shaft speed. The PSD’s of signals
taken at 2000 RPM (Figure 30) showed an increase in power throughout the entire
spectrum compared to data taken at 1000 RPM. The PSD’s also showed that the
fundamental mesh frequency, previously not apparent at 1000 RPM, became more
obvious in the PSD’s at 2000 RPM. All other peaks associated with higher order gear
meshing frequencies were not as apparent as they had been at slower shaft speeds.
The distributions of the damage feature taken at 2000 RPM for worn tooth
damage (Figure 31) showed increased variability and outliers but also an increased
separation in the means of the damaged and undamaged states. Most of the damage
features had improved performance compared to features taken at1000 RPM. The ROC
curves (Figure 32) showed that the top three performers were the crest factor, FM0 and
FMH. It was observed that the damage features that were extracted from the difference
70
signal all had swapped hypothesis from when they were extracted at the slower shaft
speed.
Figure 30: Trial Power Spectrums for Worn Tooth Damage @ 2000RPM, 2048 Samples per Revolution, Channel 1
-40 -30 -20 -100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Mean Amplitude (dB)Instances
Fre
quen
cyPower Spectral Density
Baseline Condition
50 100 150 200 2500
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityWorntooth Condition
50 100 150 200 2500
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
-40
-35
-30
-25
-20
-15
-10
71
Figure 31: Damage Feature Distributions for Worn Tooth Gearbox Damage @ 2000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1
Figure 32: Damage Feature Statistical Performance for Worn Tooth Gear Box Damage Detection @ 2000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
4
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6AM8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
72
3.4.2. Main Shaft Ball Bearing Damage Preliminary Results
The next damage case to be examined was that of bearing damage. In the
preliminary bearing damage examination two types of damage were examined under
two different conditions. Damage features for ball fault damage were first examined
under normal operating conditions then the belt drive tension was increase slightly to
see how this would affect the detection performance of the damage features. After that a
damage case of outer race bearing damage was explored as well as the effect of the
position of the fault relative to the loading of the belt drive. For the extraction of the
damage features, the angular resampling rate was always kept constant at 2048 samples
per second and a total of 20 revolutions used. From experience it was determined that
channel three was the best channel for detecting bearing damage using the damage
features. Channel 3 was the side mounted bearing, which was in the path of the load
vector from the belt drive.
When first looking at the power spectral densities of the various ball fault
damage trials (Figure 33), groups of sporadic vertical lines spanning the entire range of
frequencies can clearly be seen. These were representative of long periods of ball fault
excitation that occurred within an instance of data collection. During machine operation
while the ball fault bearing was in place, periods of low fault excitation would occur
where the ball fault would not be active. The cause of these periods of low damage
excitation was believed to be that the fault had slipped to the side and was not in the
path of any loading. The damage excitation was observed to be somewhat of a random
event and not always included in the acquired signal. The use of a longer signal
sampling length was more advantageous for use in damage detection of ball fault
damage in order to record periods of excitation. This property of ball fault damage
73
dictated the use of the maximum signal length taken which consisted of 20 shaft
revolutions or roughly 5 seconds of data.
When looking at the damage features performance, all pick up on this damage
very well. The three worst performers were FMH, NB4* and the root mean square. The
null hypothesis for all the damage features all indicate that damage is present when the
damage features showed an increase from their baseline values (Figure 35). Looking
back at the worn tooth gearbox damage case some of the null hypotheses for the
damage features indicated a reduction in value from their baseline means. When
looking at the variances in the damage feature values (Figure 34), the range between the
25 and 75th percentile values has shown a large increase compared to variance in feature
values for worn tooth damage.
Figure 33: Trial Power Spectra for Main Shaft Bearing Ball Fault Damage @ 1000RPM, 2048 Samples per Revolution, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBall Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
74
Figure 34: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
Figure 35: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH0
1
2
3
4
5
6
7
8
9
10
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMSFM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
75
The high belt tension tests produced a noticeable increase in the spectral power
from the majority of the frequency domain (Figure 36). There was limited overlap in the
feature distributions (Figure 37) which resulted in increased detection performance had
for all detection features (Figure 38) except for the FMH which was the lowest
performing damage feature. FMH’s poor performance occurs because it is tuned to
detect impacts at the meshing frequencies of orders 1-3, it was the least likely to detect
this class of damage, whereas the FM0 feature, also tuned to specific frequency orders
associated with gear mesh signals is only monitoring for changes in a certain frequency
bin. If impacts due to bearing faults are causing wide band frequency response then it
will still register this increase in energy as damage. Though this isn’t an issue for
detecting damage, it does mean that its value as a good classifying statistic is limited, as
it can’t distinguish bearing damage from gear damage.
Figure 36: Trial Power Spectrums for Main Shaft Bearing Ball Fault Damage under High Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBall Fault Condition
50 100 1500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
76
Figure 37: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage Under High Belt Tension @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
Figure 38: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Under High Belt Tension Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH0
1
2
3
4
5
6
7
8
9
10
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMSFM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
77
In the case of outer race fault bearing damage the location of the bearing fault
was marked on the side of the bearing casing with a little mark and the word outer was
engraved on the outside of the bearing to clearly distinguish it from the other bearings.
For the first test of outer race bearing damage the location of the bearing damage was
located in the same position every time. This position was closet to channel 2 and also
would have the highest normal force applied to it as it was in line with the force vector
associated with the belt tension.
When examining the power spectral densities of the outer race bearing damage
(Figure 39), it is difficult to tell if there is any damage present at all. There was less
environmental variability from changing the machine components around than there
was in the case of the baseline condition. The damage feature distributions (Figure 40)
show a great deal less variance than the ball fault bearing damage feature distributions.
These results appear to indicate that there is less damage in the damaged data than the
undamaged data. When looking at the hypotheses (Figure 41), all but one damage
feature hypotheses indicate that damage is being present when the value of the damage
feature is less than the threshold. The best performers here were NA4*, RMS, and
FMH.
78
Figure 39: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage @ 1000RPM, 2048 Samples per Revolution, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityOutter Race Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
79
Figure 40: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
Figure 41: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage Detection @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
1
1.5
2
2.5
3
3.5
4
4.5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
80
Data were taken for outer race fault bearing damage where the outer race fault
was placed at random positions with no particular order. This random positioning of the
outer race fault showed and increased variability in the PSD’s (Figure 42) when
compared to the previous outer race fault data with constant fault position. The variance
in the damage features remained small (Figure 43) compared to the ball bearing fault
damage feature values. This observed feature variance was consistent to the previous
outer race fault damage feature distributions. Hypotheses made for optimal damage
detection (Figure 44) indicated damage as being present when the value of the damage
feature is less than the threshold for all but one damage feature. The best performing
damage features were NA4*, RMS, and FMH. The FMH damage feature detected
damage in the outer race fault even though it was designed for worn tooth damage
detection. This result was unexpected because FMH is a measure of the periodicity of
impacts associated with the gear mesh harmonics and is unknown why FMH is a top
performer for this class of damage.
81
Figure 42: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage w/ Varying Fault Position @ 1000RPM, 2048 Samples per Revolution, Channel 3
Figure 43: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage w/ Varying Fault Position @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)InstancesFre
quen
cy
Power Spectral DensityOutter Race Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
1
1.5
2
2.5
3
3.5
4
4.5
5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
82
Figure 44: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage Detection w/ Varying Fault Position @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions,
Channel 3
3.4.3. Bearing Damage Detection and Varying Belt Tension
For the following results, it was desirable to try and reducing some of the
variability associated with time and manipulation of the testbed caused by replacing
parts. The effect of belt tension on bearing damage detection was to be examined in this
section and testing procedure two was used where features were taken over a single
long run with automated data collection. The results showed that the detection
performance was greatly increased for many bearing damage types even though the belt
tension was lowered by half that of the optimal belt tension for half the trials. For all the
cases presented in this section, low level and high level belt tension are represented on
the left and right half of the PSD’s respectively.
For the ball fault damage case, by looking at the PSD of the baseline condition
and the damaged condition (Figure 45), it is possible to see that there is an increase in
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
83
power to frequency components that were present in the low belt tension case. The ball
fault bearing damage still exhibits the random events of increased broadband frequency
energy that was present in the earlier case but the increase of belt tension seems to only
increase the power of these events.
When looking at the distributions for the damage features (Figure 46), it can be
seen that the performance is exceedingly well for all the damage features. The mean of
all the damage features increased for the ball fault damage case as well the base line
case had very little variability in the features when compared to the damage case.
Because the features were extracted from data sets where there was less manipulation to
the system as previous data sets, this is believed to be the reason for the drastic increase
in performance of all damage features when compared to the previous results. The ROC
curves (Figure 47) showed near perfect detection for all damage features.
84
Figure 45: Trial Power Spectrums for Main Shaft Bearing Ball Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityOutter Race Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
85
Figure 46: Damage Feature Distributions for Main Shaft Bearing Ball Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
Figure 47: Damage Feature Statistical Performance for Main Shaft Bearing Ball Fault Damage Detection w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 20 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
2
4
6
8
10
12
14
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8A NA4* NB4* FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
86
In the inner race fault bearing damage case the PSD of the trials (Figure 48)
again show that increasing the belt tension amplified the frequency content that was
originally present in low tension case. For the inner race fault, by comparing the PSD’s
of the trial it can be seen that the there is a large increase in power in the band width of
3-7 orders of the gear mesh fundamental frequency. This increase in power dominates
the gear mesh frequencies in that region, which were clearly present in the undamaged
case. Because the damage on the inner race is static and cannot move around, the
random events of broad band energy spikes for the ball bearing damage are no longer
present. Instead it can be seen that there is a uniform increase in broad band power due
to the inner race faults.
The distributions of feature values for the inner race fault (Figure 49)
continually showed an increase for the damaged case for all feature values. The damage
feature variances were again much larger than those for the baseline condition. The
performance of the damage features (Figure 50) was essentially the same as it was for
the ball fault damage. All features yielded superior detection capabilities for the inner
race ball fault damage under varying belt tension. This increase in performance is most
likely due to the reduced environmental variability that was attained by using testing
procedure 2.
87
Figure 48: Trial Power Spectrums for Main Shaft Bearing Inner Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3
Figure 49: Damage Feature Distributions for Main Shaft Bearing Inner Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)InstancesFre
quen
cy
Power Spectral DensityInner Race Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
88
Figure 50: Damage Feature Statistical Performance for Main Shaft Bearing Inner Race Fault Damage Detection w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel
3
Lastly, the outer race bearing fault was examined using testing procedure 2 for
varying belt tensions. The variability in the PSD’s (Figure 51) for this comparison
showed very little difference between the damaged and undamaged states. In the
damage case there is a slight increase in high frequency power in the bandwidth of 20-
25 orders of the gear mesh fundamental frequency. Though in the baseline condition
there is a slight increase of power in the bandwidth of 19-12 orders of the fundamental
gear mesh frequency for the results taken with the manufacturers recommended belt
tension. These results strongly suggest that the damage seeded in the outer race fault
was very small when compared to the ball fault and inner race fault damage cases. An
inquiry was made to SpectraQuest about the procedures used in seeding damage for the
different bearing faults but there has been no response as of this time.
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
89
Figure 51: Trial Power Spectrums for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, Channel 3
When examining the performance of the damage features for the outer race fault
under varying belt tension (Figure 52), the large increase in performance that was seen
in the previous two bearing damage cases is not observed. This observed result is most
likely due to the small size of the outer race fault. The ROC plots (Figure 53) showed
that the top three performers were FM0, NB4* and FMH. FM0 and FMH were tuned to
track frequency content for the first three harmonic orders. It is uncertain why they
would be the top performers for a bearing damage case taken from a sensor that
positioned on the bearing housing. Clearly the bearing fault causes a change in
frequency content that is primarily associated with the gear mesh harmonics.
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityOutter Race Fault Condition
50 1001502002500123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
90
Figure 52: Damage Feature Distributions for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3
Figure 53: Damage Feature Statistical Performance for Main Shaft Bearing Outer Race Fault Damage w/ Varying Belt Tension @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
91
Lastly, one other interesting observation made here is that the ROC plots for the
root mean square and NA4* show a distinct step pattern. This result is due to a bimodal
distribution in the damage features and indicates that the root mean square of the system
and the NA4* damage features have a clear correlation with the belt tension.
Over all reducing the environmental variability showed a large increase in
performance for all damage features in the case of the ball fault and inner race bearing
damage conditions. This result of increased detection performance was not observed for
the outer race bearing damage condition. The belt tension and normal force clearly has
an effect on the power of the frequency content but not enough to affect the
performance of the damage features in most cases, the exception being the RMS and
NA4* damage features for outer race fault condition which is believed to be minimal in
size compared the ball fault and inner race faults.
3.4.4. Worn tooth Damage and Varying Torsional Load
In this section testing procedure 2 was used to reduce environmental variability
and isolate variability caused by varying torsional load applied to the gear box. The data
sets used were ordered in increasing torsional load and can be view from left to right in
the PSD plots below. 128 trials were taken at three levels of torsional load. To
determine the effect that sensor placement has on worn tooth damage detection a
comparison was made between Channel 1 and Channel 3 data. Channel 1 data were
taken for a sensor directly mounted on the gear box while Channel 3 data were taken
from a sensor mounted to the bearing housing.
The PSD’s for Channel 1 data (Figure 54) showed similar results from the
preliminary worn tooth damage case. The undamaged case showed more power in the
high frequency range caused by the impacting of healthy gear teeth and reduced high
92
frequency content due to the slipping of worn teeth for the damaged case. When
comparing the low and high torsional loads, the power spectrums show that there is a
definite increase in power for frequencies associated with gear mesh harmonics under
higher torsional loads.
The distributions of feature values using Channel 1 data for undamaged and
damaged states (Figure 55) showed an increase in separation and reduced variance
compared to the worn tooth damage feature distributions in the preliminary results. The
ROC curves of the damage features (Figure 56) performance showed an overall increase
compared to the preliminary results. This observed increase in performance is attributed
to the reduced environmental variability achieved by using testing procedure 2. The
variability of the torsional load did not reduce the performance of the various damage
features.
93
Figure 54: Trial Power Spectrums for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, Channel 1
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
100 200 3000123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityWorn Tooth Condition
100 200 3000123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
94
Figure 55: Damage Feature Distributions for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1
Figure 56: Damage Feature Statistical Performance for Worn Tooth Damage w/ Varying Torsional Load @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 1
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
2
4
6
8
10
12
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
95
Data from Channel 3 showed presence of the gear mesh harmonics in their
PSD’s (Figure 57) though not as defined as the data taken from the sensor directly
mounted to the gearbox. The reduction of power to higher ordered frequency content
due to worn tooth slippage is present in the Channel 3 data again. The effect of the
increase in torsional load on PSD’s using Channel 3 data had the same result as the gear
box mounted sensor data. This observed result was that the frequency power associated
with the gear mesh harmonics is increased as the torsional load increases.
Figure 57: Trial Power Spectrums for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, Channel 3
The distributions of the damage feature values (Figure 58) showed more overlap
between damaged and undamaged feature distribution than was observed for feature
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityBaseline Condition
100 200 3000123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
-40 -30 -20 -100123456789
10111213141516171819202122232425262728293031323334353637
Mean Amplitude (dB)Instances
Fre
quen
cy
Power Spectral DensityWorn Tooth Condition
100 200 3000123456789
10111213141516171819202122232425262728293031323334353637
-40
-35
-30
-25
-20
-15
-10
96
values using data from Channel 1. The detection performance of the damage features
using the Channel 3 data (Figure 59) showed the top two performing damage features
were NB4* and FMH. The remaining damage features suffered a reduced ability to
detect the worn tooth damage caused by their proximity to the damage. NB4* and FMH
showed the most potential for monitoring worn tooth damage at greater distances from
the source.
Figure 58: Damage Feature Distributions for Worn Tooth Damage w/ Varying Torsional Load @ 1000RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3
CF KURT RMS FM0 FM4 M6A M8A NA4* NB4* FMH
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
Damage Feature
Damage Feature Distribution
Dam
ages
Med
ian
Nom
aliz
ed F
eatu
re V
alue
97
Figure 59: Damage Feature Statistical Performance for Worn Tooth Damage w/ Varying Torsional Load @ 1000 RPM, 2048 Samples per Revolution, 40 Revolutions, Channel 3
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Damage Feature ROC Plot
Prob. of False Alarm: PFA
Pro
b. o
f D
etec
tion:
PD
CFKURT
RMS
FM0
FM4M6A
M8A
NA4*
NB4*FMH
CF KURTRMS FM0 FM4 M6A M8ANA4*NB4*FMH0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Damage Feature
Tot
al P
roba
bilit
y (%
)
Total Probability: Area Under ROC
98
Chapter 4
Conclusion & Future Work
4.1. Conclusion
From the results of this study it was shown that the relation of the performance
of various damage features to environmental variability and operational conditions is
complex. Many of the damage features indicated good performance for the different
damage types under different condition but suffered from hypothesis swapping, where
for one system state, damage caused an increase in the mean of the feature value, but in
others the damage caused a decrease. The change in damage feature distributions makes
it hard to establish specified thresholds to detect if damage is present. Classification of
damage types become increasingly also become difficult.
Out of the damage features covered, the newly proposed FMH feature
continuously showed superior performance in detecting worn tooth damage compared
to traditional gearbox damage features. This superior performance occurred for a
number of different signal lengths, sampling frequencies and sensor locations. In
addition to its superior performance in worn tooth damage detection, the FMH feature
also was less sensitive to bearing damage than the traditional damage features. Lastly,
99
because damage nearly always produced an increase in the feature value of the FMH
damage feature the hypotheses made for optimal detection was rarely swapped unlike
the traditional damage features.
Two main contributions have been made in this thesis. The first contribution
was providing a better understanding of traditional damage features performance under
varying environmental and operational conditions. The survey of the damage detection
performance for various damage states can be extrapolated for use in industry. The
knowledge of which features perform better for certain classes of damage and don’t
suffer from swapping hypotheses can be used for more informed damage classification
of damage types. This will allow CBM system to achieve better reliability and
performance. The second and most fruitful contribution was the introduction of the
FMH damage feature for use in detecting worn tooth damage. The results in this thesis
showed the FMH damage to be an excellent candidate for detecting early stages of gear
damage. With earlier detection of gear wear, better planning of maintenance and
ordering of replacement parts can be achieved. This results in a reduction of cost and
increase in potential safety.
4.2. Recommendations for Future Work
This study has provided valuable information on the effect of varying
environmental and operational parameters of various damage features under a number
of damage conditions that occur in rotating machines. The majority of the damage
features performed well under certain conditions while others did not. As such they can
be used for group classification of various damage types as well as location in
multidimensional data driven analysis.
100
The next step after the work presented in this study would be for the damage
features to be compared in a statistically significant way by using groups of features
instead of a single one then compared against which type of damage was present. The
multi-dimensional statistics of a group of features for each damage type can be learned
and then machine damage can be classified using the Mahalanoblis distance of the
damage feature from the trained statistics in an attempt to be able to classify the type of
damage present.
101
References
[1] C. R. Farrar and K. Worden, “An introduction to structural health monitoring,” Phil. Trans. R. Soc. A, vol. 365, no. 1851, pp. 303–315, Feb. 2007.
[2] C. R. Farrar and T. A. Duffey, “VIBRATION-BASED DAMAGE DETECTION IN ROTATING
MACHINERY,” in The International Conference on Damage Assessment of Structures (DAMAS 99), Dublin, Ireland, 1999.
[3] C. R. Farrar, S. W. Doebling, and D. A. Nix, “Vibration–based structural damage identification,”
Phil. Trans. R. Soc. Lond. A, vol. 359, no. 1778, pp. 131–149, Jan. 2001. [4] M. Mosher, A. H. Pryor, and E. M. Huff, “Evaluation of Standard Gear Metrics in Helicopter Flight
Operation," presented at 56th Meeting of the Society for Machinery Failure Prevention Technology.” [5] R. B. Randall, “Detection and diagnosis of incipient bearing failure in helicopter gearboxes,”
Engineering Failure Analysis, vol. 11, no. 2, pp. 177–190, Apr. 2004. [6] P. A. Higgs, R. Parkin, M. Jackson, A. Al-Habaibeh, F. Zorriassatine, and J. Coy, “A survey on
condition monitoring systems in industry,” in The 7th Biennial ASME Conference Engineering Systems Design and Analysis, 2004, pp. 19–22.
[7] F. P. García Márquez, A. M. Tobias, J. M. Pinar Pérez, and M. Papaelias, “Condition monitoring of
wind turbines: Techniques and methods,” Renewable Energy, vol. 46, pp. 169–178, Oct. 2012. [8] Z. Hameed, Y. S. Hong, Y. M. Cho, S. H. Ahn, and C. K. Song, “Condition monitoring and fault
detection of wind turbines and related algorithms: A review,” Renewable and Sustainable Energy Reviews, vol. 13, no. 1, pp. 1–39, Jan. 2009.
[9] W. Lu and F. Chu, “Condition monitoring and fault diagnostics of wind turbines,” in Prognostics
and Health Management Conference, 2010. PHM ’10., 2010, pp. 1–11. [10] R. F. Orsagh, H. Lee, M. Watson, C. Byington, and J. Powers, “Advanced Vibration Monitoring for
Wind Turbine Health Management.” .
102
[11] L. F. Villa, A. Reñones, J. R. Perán, and L. J. de Miguel, “Angular resampling for vibration analysis in wind turbines under non-linear speed fluctuation,” Mechanical Systems and Signal Processing, vol. 25, no. 6, pp. 2157–2168, Aug. 2011.
[12] B. Al-Najjar and I. Alsyouf, “Enhancing a company’s profitability and competitiveness using
integrated vibration-based maintenance: A case study,” European Journal of Operational Research, vol. 157, no. 3, pp. 643–657, Sep. 2004.
[13] B. Al-Najjar, “The lack of maintenance and not maintenance which costs: A model to describe and
quantify the impact of vibration-based maintenance on company’s business,” International Journal of Production Economics, vol. 107, no. 1, pp. 260–273, May 2007.
[14] F. Barbera, H. Schneider, and P. Kelle, “A Condition Based Maintenance Model with Exponential
Failures and Fixed Inspection Intervals,” The Journal of the Operational Research Society, vol. 47, no. 8, pp. 1037–1045, Aug. 1996.
[15] T. Biagetti and E. Sciubba, “A First Step Towards Unmanned Intelligent Process Management: A
Procedure for the Diagnostics and Prognostics of Energy Conversion Plants,” International Journal of Thermodynamics, vol. 5, no. 2, pp. 85–99, Feb. 2010.
[16] P. J. Vlok, J. L. Coetzee, D. Banjevic, A. K. S. Jardine, and V. Makis, “Optimal Component
Replacement Decisions Using Vibration Monitoring and the Proportional-Hazards Model,” The Journal of the Operational Research Society, vol. 53, no. 2, pp. 193–202, Feb. 2002.
[17] R. B. Randall, Vibration-based condition monitoring: industrial, aerospace, and automotive
applications. Chichester, West Sussex, U.K. ; Hoboken, N.J: Wiley, 2011. [18] J. J. Zakrajsek, D. P. Townsend, and H. J. Decker, “An Analysis of Gear Fault Detection Methods as
Applied to Pitting Fatigue Failure Data,” Jan. 1993. [19] H. J. Decker and J. J. Zakrajsek, “Comparison of Interpolation Methods as Applied to Time
Synchronous Averaging,” Jun. 1999. [20] H. J. Deckeer, R. F. Handschuh, and J. J. Zakrajsek, “An Enhancement to the NA4 Gear Vibration
Diagnostic Parameter.,” Jul. 1994. [21] J. J. Zakrajsek, “An Investigation of Gear Mesh Failure Prediction Techniques,” Oct. 1989. [22] J. J. Zakrajsek, D. P. Townsend, D. G. Lewicki, H. J. Decker, and R. F. Handschuh, “Transmission
Diagnostic Research at NASA Lewis Research Center.,” Nov. 1995. [23] S. T. Lin and P. D. McFadden, “GEAR VIBRATION ANALYSIS BY B-SPLINE WAVELET-
BASED LINEAR WAVELET TRANSFORM,” Mechanical Systems and Signal Processing, vol. 11, no. 4, pp. 603–609, Jul. 1997.
[24] Z. K. Peng and F. L. Chu, “Application of the wavelet transform in machine condition monitoring
and fault diagnostics: a review with bibliography,” Mechanical Systems and Signal Processing, vol. 18, no. 2, pp. 199–221, Mar. 2004.
[25] P. D. McFadden and W. Wang, “ORA Report: ‘Time-frequency domain analysis of vibration signals
for machinery diagnostics (1) Introduction to the Wigner-Ville distribution’ - uuid:f9025874-c208-4984-8935-cb3801b8a04a.”[Online]. Available: http://ora.ox.ac.uk/objects/uuid:f9025874-c208-4984-8935-cb3801b8a04a. [Accessed: 13-Aug-2013].
103
[26] W. J. Staszewski, K. Worden, and G. R. Tomlinson, “TIME–FREQUENCY ANALYSIS IN GEARBOX FAULT DETECTION USING THE WIGNER–VILLE DISTRIBUTION AND PATTERN RECOGNITION,” Mechanical Systems and Signal Processing, vol. 11, no. 5, pp. 673–692, Sep. 1997.
[27] X. Deng, Q. Wang, and X. Chen, “A Time-Frequency Localization Method for Singular Signal
Detection Using Wavelet-Based H #x0F6;lder Exponent and Hilbert Transform,” in Congress on Image and Signal Processing, 2008. CISP ’08, 2008, vol. 4, pp. 266–270.
[28] A. N. Robertson, C. R. Farrar, and H. Sohn, “Singularity detection for structural health monitoring
using Holder exponents,” 2003, vol. 5057, pp. 569–581. [29] N. B. Do, S. R. Green, and imothy A. Schwartz, “Structural Damage Detection Using the Holder
Exponent,” in The 21st IMAC Conference on Structural Dynamics, 2003. [30] T. Nakamura, H. Horio, and Y. Chiba, “Local Holder exponent analysis of heart rate variability in
preterm infants,” IEEE Transactions on Biomedical Engineering, vol. 53, no. 1, pp. 83–88, 2006. [31] J. Antoni, “Fast computation of the kurtogram for the detection of transient faults,” Mechanical
Systems and Signal Processing, vol. 21, no. 1, pp. 108–124, Jan. 2007. [32] J. Antoni and R. B. Randall, “The spectral kurtosis: application to the vibratory surveillance and
diagnostics of rotating machines,” Mechanical Systems and Signal Processing, vol. 20, no. 2, pp. 308–331, Feb. 2006.
[33] J. Antoni, “The spectral kurtosis: a useful tool for characterising non-stationary signals,” Mechanical
Systems and Signal Processing, vol. 20, no. 2, pp. 282–307, Feb. 2006. [34] N. Sawalhi and R. B. Randall, “Spectral kurtosis optimization for rolling element bearings,” in
Proceedings of the Eighth International Symposium on Signal Processing and Its Applications, 2005, 2005, vol. 2, pp. 839–842.
[35] E. Bechhoefer, “A Quick Introduction to Bearing Envelope Analysis.” [Online]. Available:
http://www.scribd.com/doc/110769949/A-Quick-Introduction-to-Bearing-Envelope-Analysis. [Accessed: 12-Aug-2013].
[36] A. Heng, S. Zhang, A. C. C. Tan, and J. Mathew, “Rotating machinery prognostics: State of the art,
challenges and opportunities,” Mechanical Systems and Signal Processing, vol. 23, no. 3, pp. 724–739, Apr. 2009.
[37] W. Wang, P. A. Scarf, and M. A. J. Smith, “On the Application of a Model of Condition-Based
Maintenance,” The Journal of the Operational Research Society, vol. 51, no. 11, pp. 1218–1227, Nov. 2000.
[38] J. Yan, “A prognostic algorithm for machine performance assessment and its application,”
Production Planning & Control: The Management of Operations,, vol. 15, pp. 796–801, 2004. [39] E. B. Flynn and M. D. Todd, “A Bayesian approach to optimal sensor placement for structural health
monitoring with application to active sensing,” Mechanical Systems and Signal Processing, vol. 24, no. 4, pp. 891–903, May 2010.
[40] Q. Liu and H. Wang, “A case study on multisensor data fusion for imbalance diagnosis of rotating
machinery,” Artif. Intell. Eng. Des. Anal. Manuf., vol. 15, no. 3, pp. 203–210, Jun. 2001.
104
[41] P. D. Welch, “The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms,” IEEE Transactions on Audio and Electroacoustics, vol. 15, no. 2, pp. 70–73, 1967.
[42] J. Markel, “Digital inverse filtering-a new tool for formant trajectory estimation,” IEEE Transactions
on Audio and Electroacoustics, vol. 20, no. 2, pp. 129–137, 1972. [43] “Mother Wavelets.” [Online]. Available: http://www.ljmu.ac.uk/GERI/98293.htm. [Accessed: 22-
Aug-2013]. [44] F. Bonnardot, M. El Badaoui, R. B. Randall, J. Danière, and F. Guillet, “Use of the acceleration
signal of a gearbox in order to perform angular resampling (with limited speed fluctuation),” Mechanical Systems and Signal Processing, vol. 19, no. 4, pp. 766–785, Jul. 2005.
[45] J. R. Blough, “Adaptive Resampling-Transforming from the Time to the Angle Domain,” IMAC-
XXIV: Conference & Exposition on Structural Dynamics, 2006. [46] E. Bechhoefer and M. Kingsley, “A review of time synchronous average algorithms,” Annual
Conference of the Prognostics and Health Management Society, 2009. [47] F. Combet and L. Gelman, “An automated methodology for performing time synchronous averaging
of a gearbox signal without speed sensor,” Mechanical Systems and Signal Processing, vol. 21, no. 6, pp. 2590–2606, Aug. 2007.
[48] K. McClintic, M. Lebold, K. Maynard, C. Byington, and R. Campbell, “Residual and difference
feature analysis with transitional gearbox data,” In 54 th Meeting of the Society for Machinery Failure Prevention Technology, 2000.
[49] P. D. McFadden, “Interpolation techniques for time domain averaging of gear vibration,”
Mechanical Systems and Signal Processing, vol. 3, no. 1, pp. 87–97, Jan. 1989. [50] J. Antoni and R. B. Randall, “Unsupervised noise cancellation for vibration signals: part I—
evaluation of adaptive algorithms,” Mechanical Systems and Signal Processing, vol. 18, no. 1, pp. 89–101, Jan. 2004.
[51] J. Antoni and R. B. Randall, “Unsupervised noise cancellation for vibration signals: part II—a novel
frequency-domain algorithm,” Mechanical Systems and Signal Processing, vol. 18, no. 1, pp. 103–117, Jan. 2004.
[52] M. K. P. Vev cev r, “Condition Indicators for Gearbox Condition Monitoring Systems,” Acta
Polytechnica, vol. 45, pp. 35–43, 2005. [53] A. S. Sait and Y. I. Sharaf-Eldeen, “A Review of Gearbox Condition Monitoring Based on vibration
Analysis Techniques Diagnostics and Prognostics,” in Rotating Machinery, Structural Health Monitoring, Shock and Vibration, Volume 5, T. Proulx, Ed. Springer New York, 2011, pp. 307–324.
[54] M. Lebold, K. McClintic, R. Campbell, C. Byington, and K. Maynard, “Review of vibration analysis
methods for gearbox diagnostics and prognostics.” 2000. [55] A. P. Bradley, “The use of the area under the ROC curve in the evaluation of machine learning
algorithms,” Pattern Recognition, vol. 30, no. 7, pp. 1145–1159, Jul. 1997.